Education
Concept Leap
Concept Leap is an AI learning companion for students doing self-study and getting stuck without real-time help. Instead of only answering questions, it aims to identify the concept, explain it at the right level, check understanding, and adapt the teaching path when a learner is still confused. The demo centers on concept mastery as the product goal, with analogies, quick checks, deeper explanations, and feedback loops built into a single interaction.
The problem
A CBSE Grade 6 student sits with an NCERT Science textbook at 7 PM, stuck on a paragraph, with no one able to help. Re-reading only deepens the confusion. The options are all poor: a parent who can't remember the material, a 15-minute YouTube video where the needed 30 seconds is hard to find, or general-purpose chatbots that give college-level or oversimplified answers that don't match the textbook. The student typically gives up, moves on with incomplete understanding, and gaps accumulate into poor test performance. The most severe pains are "I'm stuck on this paragraph right now," answers that don't match their level, and getting an answer without understanding why.
The solution
ConceptLeap is an AI learning companion for Grade 6 CBSE students during self-study. A student photographs a confusing textbook paragraph or types a question and gets an instant, grade-appropriate explanation — with the goal of concept mastery, not just an answer. Its differentiators are CBSE-native, grade-aware explanations framed in NCERT language; image-to-understanding rather than image-to-answer; and a Socratic-first approach that asks a guiding question before explaining, falling back to direct explanation if the student is frustrated. Parents see a simple daily list of topics asked about — engagement signal without surveillance.
How it works
The core learning flow is Question → Analogy → Quick Check (a single 3-option multiple-choice question with one correct answer) → Explanation → Real-World Connection → Concept Unlocked. It uses GPT-5 for its multimodal understanding of textbook pages, reliable Socratic instruction-following, age-appropriate dialogue, and support for English, Hindi, and Hinglish; Gemini 2.5 Flash and Claude Sonnet were evaluated but rejected on Socratic consistency and multimodal maturity respectively. The production system prompt (v5.0) restricts scope to Grade 6 Science and Mathematics, keeps responses under 150 words at a Grade 6 reading level, uses positive reinforcement before correction, and refuses unsafe or out-of-scope requests. The MVP uses prompt engineering rather than production RAG, though NCERT-based retrieval is planned. Across 30 test scenarios, the core tutoring workflow performed successfully on most cases; identified gaps include diagram-only images, ambiguous chapter references, and prompt-injection resistance.
Who it's for
The product is B2C with a clear buyer-user split. The daily user is a CBSE Grade 6 student aged 11–12 who struggles with NCERT paragraphs during 6–9 PM self-study and already photographs textbook pages naturally. The buyer is a working parent, typically 35–45 and dual-income, who can't help with Grade 6 Science and wants something affordable (under ₹500/month), safe, effective, and usable without supervision. Revenue is a freemium subscription — basic concept explanations free, with premium proficiency tracking, personalized practice, and parent analytics — with future school partnerships and exam-prep modules.
Why it matters
The Indian EdTech market is projected to grow from roughly USD 3.6–7.5B in 2025 toward USD 29–33B by 2030–2034 at about 28% CAGR, and India's AI-in-education market from USD 196M (2024) to over USD 1.1B by 2030 at ~32% CAGR, with 250M+ CBSE students underserved by after-school tutoring. An early-stage MVP, ConceptLeap plans a phased rollout starting with a 4-week pilot of ~50 Grade 6 students. Its North Star is a Concept Mastery Rate above 70% of sessions reaching Concept Unlocked, with supporting targets like a "This Helped" rate above 80%. Post-2022 EdTech trust issues and the non-negotiable accuracy bar for K-12 make grounded, guardrailed tutoring essential.
The workflow
The PRD
| PRODUCT FACULTY — AI PRODUCT REQUIREMENTS DOCUMENT (PRD) TEMPLATE Version 1.0 | ||||||
|---|---|---|---|---|---|---|
| Your Name: | Srikanth R Rao K | |||||
| Your Product: | ConceptLeap | |||||
| Your Industry: | Education | |||||
| Date: | May 21, 2026 | |||||
| 4D Method | AI PRD | Instructor Feedback | ||||
| Phase | Activity | Theme | Topic | Key Question(s) | Your Response Include external links to visuals/prototypes as required. | |
| DISCOVERY | Understand your market, business, product & user context | Business Value Map | Market Attractiveness | What industry is your business in? (ie Financial services, Healthcare, Education, etc)? | EdTech - specifically AI-powered K-12 tutoring for the Indian market. ConceptLeap operates at the intersection of Generative AI and after-school self-study support for CBSE students. | Srikanth, the buyer-user separation is the right structural foundation. Mapping the parent as purchaser and the student as daily user is a decision that shapes everything downstream, and you made it early. Three gaps, in priority order. First, AI necessity. Every capability listed in Discovery (concept simplification at multiple depths, practice generation, mastery scoring) could be delivered by a curated content library with a rules engine. The LLM earns its place only when the system must reason over unstructured student input in real time, like a photographed paragraph paired with a follow-up question in the student's own words. Isolate those moments and cut the rest from the AI hypothesis. Second, the journey map documents the product experience, not the student's world before the product exists. Rebuild it around what a struggling student actually does today: searching YouTube, waiting for tuition, asking a parent who may not remember the material. Without that baseline, there is no friction to measure against. Third, the Design section names three parallel AI workflows but does not pick one as the MVP. One workflow, one narrow grade band, one set of evaluation criteria is the scope that makes the next phase buildable. The competitive landscape also needs a level of depth beyond category labels. For each named player, document the specific capability it handles well and the gap it leaves open. That is what makes positioning defensible on Demo Day rather than asserted. Pick the single workflow closest to the pain, define what good output looks like for it, and write the initial prompt. That is the path from here to a testable prototype. |
| What are the key challenges (headwinds) and opportunities (tailwinds) impacting growth in your industry? Who are the key competitors? | Tailwinds: 1) Indian EdTech market projected to grow from USD 3.6-7.5B (2025) to USD 29-33B (2030-2034) at 28% CAGR. 2) India's AI-in-Education market: USD 196M (2024) → USD 1.1B (2030) at 32% CAGR. 3) 250M+ CBSE students with limited access to quality after-school tutoring Headwinds: 1) Post-2022 EdTech trust deficit (BYJU'S crisis, funding winter). 2) Price sensitivity in Indian middle-class families. 3) Parents skeptical of "screen time = learning" claims. 4) Regulatory uncertainty: India's Digital Personal Data Protection Act (DPDP) for minors. 5) AI hallucination risk in educational contexts — accuracy is non-negotiable for K-12 IC's 1) BYJU'S - Video lessons; India-focused ; brand recognition. GAPS - Passive video consumption; not interactive; no real-time doubt resolution; trust issues post-2022 2) Vedantu - Live tutoring; CBSE-aligned.GAPS: Requires scheduling; expensive (₹15K-50K/year); Doubt learning not available post 7 PM, does not address Grade 6 scope 3) PhysicsWallah: Affordable; CBSE-focused; strong for Grades 11-12. GAPS - Primarily video + live class model; limited AI interaction; weaker for Grades 6 4) Doubtnut/Question AI - Answer-focused (not understanding-focused); no Socratic approach; video solutions not personalized to grade level | |||||
| What is the projected growth rate of your target market segment over the next 3-5 years? | ConceptLeap operates within the rapidly growing Indian K-12 EdTech and AI-in-Education market. 1) The Indian EdTech market was valued at approximately USD 3.6B–7.5B in 2025 and is projected to grow to nearly USD 29B–33B by 2030–2034, growing at approximately 28% CAGR. 2) India’s AI-in-Education market is projected to grow from USD 196M in 2024 to over USD 1.1B by 2030, at approximately 32% CAGR. 3) The India K-12 education market is expected to grow at approximately 11–17% CAGR through 2030, driven by NEP 2020, increasing private education spend, and growing digital learning adoption. | |||||
| Business Model | What growth stage is your business currently in (e.g., startup, scale-up, mature)? | ConceptLeap is currently positioned as an early-stage startup / MVP-stage product focused on validating product-market fit and user engagement with AI-driven study companion worklows | ||||
| How does your business make money? What do they sell? What is your primary revenue model (e.g., subscription, freemium, licensing, marketplace, transactional, etc?) | ConceptLeap will operate primarily as a B2C subscription-based platform targeted at parents of CBSE students. A freemium model can be used where basic concept explanations are free while advanced proficiency tracking, personalized practice, and parent analytics are premium features. Future monetization opportunities may include school partnerships and premium exam preparation modules. | |||||
| Who is your primary customer base (B2B, B2C, B2B2C)? | B2C — with a critical buyer-user separation: 1) Buyer (pays): Parents of CBSE Grade 6 students — typically working professionals (dual-income, urban/semi-urban, English-medium) who cannot help with Science homework after work 2) User (daily): Students aged 11-12 who struggle with NCERT textbook paragraphs during self-study (typically 6-9 PM) | |||||
| Differentiators | What are the key differentiators for your company? | 1) CBSE-native, grade-aware AI: Explanations are framed in NCERT language at the exact cognitive level of Grade 6 — not generic "explain like I'm 5" or college-level 2) Image-to-understanding (not image-to-answer): Student photographs a textbook paragraph → AI guides them to understand it (not just gives the answer like Doubtnut) 3) Socratic-first pedagogy: AI asks a guiding question before explaining — building critical thinking, not dependency. Falls back to direct explanation if student is frustrated 4) After-school, self-study context: Designed for the "7 PM stuck moment" — no teacher available, parent can't help, tuition class is tomorrow 5) Parent visibility without intrusion: Parents see "Your child asked about Friction (Ch. 12) today" — not the full conversation. Signals engagement without surveillance | ||||
| Feature Value Map (IMPORTANT NOTE: This section is only relevant if you are working on enhancing an existing product. It does not apply if you are developing a new product from 0 to 1.) | Customers | Who are the customers (ie buyers) of your product? | Parents of CBSE Grade 6 students: Age: 35-45, working professionals (dual income typical) Pain: "My child is stuck on homework and I can't help — I don't remember Grade 6 Science" Buying trigger: Sees child frustrated/disengaged OR receives poor test scores Decision criteria: Affordable (<₹500/month), safe (no inappropriate content), effective (visible improvement), independent (child can use without supervision) | |||
| End Users | Who are the end-users of your product? Which users are the most revenue-generating / revenue-impacting for your company? What are their goals, roles, and context? | Primary: CBSE Grade 6 student (age 11-12) 1) Context: Sitting with NCERT Science textbook at 7 PM, stuck on a paragraph 2) Goal: Understand the concept well enough to complete homework/prepare for tomorrow's class 3) Behavior: Photographing textbook pages is second-nature (already uses phone camera for notes) Secondary user: Parent (viewing dashboard) 1) Context: Checking child's learning activity after work (9-10 PM) 2) Goal: Confirm child is studying, see what topics they're struggling with | ||||
| Current Products / Services | If you are a Product-led business: What are the core features of your product, and how do they address user needs? If you are a Service-led business: What are the key services you offer to customers? | ConceptLeap is an AI-powered study companion focused on CBSE learning workflows. Core features include: 1) Doubt input: Photo upload OR typed text question. 2) AI concept identification: Recognizes NCERT topic from image/text, maps to Grade 6 Science/ Maths curriculum. 3) Socratic response: AI asks one guiding multiple-choice question with to activate prior knowledge before explaining. 4) Grade-appropriate explanation: <150 words, 1 analogy + 1 real-world example, NCERT-framed language. 5) Fallback to direct: If student says "just tell me" → switches to direct explanation mode. 6) One follow-up turn: Student can ask "explain in simple words/ tell me more" on the on follow-up question | ||||
| User Value Map | Target Persona | Who is your AI product / feature for? (Internal users, external users, an influencer, a buyer, etc) | Primary: External end-user — CBSE Grade 6 student (age 11-12) during self-study Buyer/Influencer: Parent who pays for the subscription and monitors the parent dashboard | |||
| Journey Map (current-state) | What is the typical journey for your target persona when they are using your product / service, focusing on their ideal experience (happy path) as they interact with your product? | Student's current journey when stuck on a Science concept (7 PM self-study): 1. [READ] Opens NCERT textbook, reads paragraph on "Chemical Effects of Electric Current" ↓ 2. [STUCK] Doesn't understand a sentence — "What does 'electrolyte' mean in this context?" ↓ 3. [ATTEMPT] Re-reads 2-3 times, gets more confused ↓ 4. [SEEK HELP - Option A] Asks parent → Parent says "I'll check later" or gives wrong/vague answer [SEEK HELP - Option B] Opens YouTube → Gets 15-min video, can't find the specific 30 seconds they need [SEEK HELP - Option C] Asks ChatGPT → Gets a college-level answer or an oversimplified one that doesn't match their textbook [SEEK HELP - Option D] Waits for tuition class tomorrow → Forgets the doubt by then ↓ 5. [GIVE UP] Skips the question/exercise, moves on with incomplete understanding ↓ 6. [CONSEQUENCE] Gaps accumulate → poor test performance → parent frustration → reactive tuition enrollment | ||||
| Pain-points | Where does the user experience friction, obstacles, or unmet needs throughout the journey? Identify which pain-points are most frequent and severe? | Key pain points include: 1) "I'm stuck on THIS paragraph right now and no one can help" [HIGH] 2) ChatGPT gives me an answer I don't understand — it's too complex/too simple - No grade-awareness; no NCERT framing; no check for understandin [HIGH] 3) "I feel stupid asking this question" - Social shame in class; judgment from tutor; parent frustration 4) "I got the answer from the internet but I still don't understand WHY" | ||||
| AI Opportunities | From your list of pain points, identify those that can effectively be addressed using Generative AI. Remember, this project focuses on leveraging LLM-powered AI to solve your target persona's pain points. Rank these pain points starting with the most severe and frequently occurring first. | 1) "I'm stuck on THIS paragraph right now" — needs instant, contextual explanation - Why LLM? Infinite combinations of textbook images × student phrasing × grade level require real-time reasoning. No static FAQ or rules engine can cover this. GenAI reads the actual photographed paragraph and generates a contextual response. 2) "ChatGPT's answer doesn't match my level" — needs grade-calibrated, NCERT-framed explanation - why LLM? LLM with constrained prompting can dynamically adjust complexity, vocabulary, and analogies to Grade 6 cognitive level — something a rules-based system cannot flexibly do across all Science topics. 3) "I got the answer but don't understand WHY" — needs Socratic guidance, not just answers - LLM is needed to generate contextually relevant guiding questions ("What do you think happens if...?") based on what the student asked. This requires reasoning over the student's specific input — not a pre-scripted Q&A tree. | ||||
| Develop an AI Solution Hypothesis | AI Solution Hypothesis | Diverge | Ideate potential solutions to address your AI-solvable pain points. Focus on generating a high quantity of ideas rather than evaluating their quality at this stage. | 1) AI Doubt Resolver (Photo → Socratic Explain) — Student uploads textbook photo, AI asks a guiding question, then explains at grade level 2) AI Practice Quiz Generator — AI generates 3-5 questions after explanation to verify understanding 3) AI Study Planner — AI recommends what to study today based on weak areas and upcoming tests 4) AI Concept Connector — AI shows how current topic links to previously learned topics (knowledge graph) 5) AI Homework Checker — Student uploads completed homework, AI identifies errors and explains corrections 6) AI Parent Report Generator — Weekly AI-generated summary email to parents on child's learning patterns 7) AI Revision Scheduler — Spaced repetition system powered by AI for long-term retention 8) AI Video Summarizer — AI watches YouTube videos and extracts the 30-second clip relevant to student's specific doubt | ||
| Converge | Rank your ideated solutions based on impact and feasibility. Identify your top three AI solutions, and clearly select the one you'll focus on for your project. | 1) AI Doubt Resolver (Photo → Socratic Explain). 2) AI Practice Quiz Generator. 3) AI Parent Report (topic visibility) Selected MVP Focus: AI-Powered Doubt Resolution with Socratic Mode — A Grade 6 CBSE Science/Math student uploads a photo of a confusing textbook paragraph (or types a question) and gets an instant, Socratic-guided, grade-appropriate explanation in <8 seconds. The AI asks a guiding question first, then explains if the student is still stuck. Parents see a simple daily list of topics asked about. | ||||
| DESIGN | Define Target State Workflow | UX Flows & Wireframes Suggested Tool: Excalidraw | Workflow (future) | Assuming your product or feature works as desired, what is the target state workflow? | Here is the workflow - link https://excalidraw.com/#json=q1h8XVy6GBsH61mlMoi8c,kB0OLKBIJZ80N7_j1oelJw Assumption: Onboarding has captured Student Name + Class (Grade 6). Key Design Decisions: 1) Max 1 follow-up turn (prevents infinite loops; MVP constraint) 2) Socratic mode is the default but student can always opt for direct explanation 3) Out-of-scope requests get a graceful decline (not an error) 4) Session always logs to parent dashboard regardless of outcome 5) No login/authentication in MVP — simple onboarding only (name + class) | Srikanth, the Discovery overhaul landed well. The pre-product journey map now traces what a struggling student actually does before the product exists, and that is a meaningful structural improvement from the prior version. The four failed alternatives ground the friction in observable behavior rather than abstraction. The evaluation criteria spreadsheet shows real rigor, with fourteen dimensions, named pass thresholds, and measurement methods that mix automated checks with LLM-as-judge scoring. The example cases spreadsheet carries ten golden path scenarios and at least eight edge cases with explicit pass conditions. That is the kind of preparation that makes the Develop phase buildable. One issue cuts across both phases and needs to be resolved before anything else. The Discovery section and the entire ICP frame the product around Grade 7 and 8 CBSE Science, but the master prompt targets Grade 6, lists Grade 6 NCERT chapters, and instructs the model to decline anything outside that scope. The evaluation criteria and example cases also appear built for Grade 6. That mismatch means the prompt, the test suite, and the Discovery hypothesis are not yet describing the same product. Pick one grade level and realign everything around it. Beyond that, the competitive landscape names four players and assigns each a gap, but none of those gap claims are sourced. One user review, one published limitation, or one personal test result per competitor would move those from assertion to evidence, which is what Demo Day judges look for in the Discovery scoring dimension. |
| Build Wireframes | Wireframes | How will users navigate through your AI solution? What are the key steps and decision points? What information will be displayed at each stage? What specific UI elements are needed on each screen? How will the layout accommodate AI features? | See Below for Prototypes | |||
| Develop Prototype to showcase AI interactions | Prototype Screens Suggested Tool: lovable.dev | What aspects of the AI solution will you demonstrate in your prototype? How will the AI inputs, processing, and outputs be presented visually to users? Which features are essential for launch? What can be left for later releases? | https://drive.google.com/file/d/1x4F5DoVqVVFcAX2VOMX7G0WpenZLGJrC/view?usp=sharing | |||
| Initial Prompt Design | Master Prompt [Initial Design] | Create an initial master prompt. Consider the following: What tone or personality should the AI use? How should user input and system instructions be structured for maximum clarity? What system instruction will govern the AI's behavior? What examples might improve performance? How will you format outputs for consistency? | https://docs.google.com/document/d/1dZPqq7ityoeX5SRVazF4OrjXtn_eIWKuR5VsqaRD7vg/edit?usp=sharing | |||
| Prepare for Testing & Iteration | Evaluation Criteria & Test Plan | Evaluation Criteria | What specific quality benchmarks (e.g., clarity, relevance, tone, accuracy, SEO, hallucination avoidance) will define “good” output? | https://docs.google.com/spreadsheets/d/16BRHg06VqNMswI8Pzn-UX3uy-DWgroyBmw2lQ2V_MAM/edit?usp=sharing | ||
| Example Cases | What specific example use cases, edge cases, and negative cases should be covered by test prompts and outputs? | https://docs.google.com/spreadsheets/d/1sebJsoLoawoSrRve03iN0ETP9AnOMq39fdxQO4AFscM/edit?usp=sharing | ||||
| DEVELOP | AI Model Selection & Justification | AI Model Selection & Justification | Which AI model is best suited for your solution and why? What capabilities and limitations does it have? How will it integrate with your product? | PRIMARY MODEL: GPT-5 (OpenAI) WHY GPT-5: ConceptLeap requires a model that simultaneously excels at: (1) Multimodal understanding of textbook content — students upload NCERT textbook pages, worksheets, diagrams, handwritten notes, and mobile photos; GPT-5 reliably extracts concepts, identifies chapters, and interprets visual learning content in a single workflow. (2) Socratic instruction-following — ConceptLeap's core differentiator is guided learning through analogy and questioning rather than direct answers. GPT-5 consistently follows structured tutoring instructions such as: identify concept → generate analogy → ask a guided question → assess understanding → adapt the next response. (3) Age-appropriate educational dialogue — Grade 6–8 students require simple language, relatable examples, and progressive scaffolding; GPT-5 produces explanations calibrated to student comprehension levels while maintaining accuracy. (4) Indian classroom context — GPT-5 handles English, Hindi, and Hinglish inputs effectively, making it suitable for Indian learners who naturally mix languages while studying. CAPABILITIES LEVERAGED: • Native multimodal reasoning (image + text in a single interaction) • Concept extraction from textbook pages, diagrams, and worksheets • Structured tutoring responses (Concept → Analogy → Guided Question → Understanding Check) • Multi-turn conversational memory for adaptive tutoring • Age-appropriate explanation generation for Grade 6–8 students • Streaming responses for responsive user experience • Metadata extraction (chapter, concept, topic) for learning summaries and dashboards LIMITATIONS & MITIGATIONS: • Can be verbose by default → System prompts enforce concise, student-friendly responses and one concept at a time • May occasionally misclassify chapter/topic from ambiguous images → Confidence-based concept detection with future curriculum-grounded retrieval planned • Higher cost than lightweight models → Acceptable for MVP; future hybrid routing architecture planned for scale optimization • Higher latency than smaller models → Streaming responses provide immediate feedback while the full answer is generated • Vendor dependence → Model abstraction layer and model-agnostic prompting strategy allow future portability INTEGRATION ARCHITECTURE: Student Question/Image → GPT-5 (Vision + Concept Identification) → GPT-5 with ConceptLeap Socratic Master Prompt → Analogy Generation → Guided Question → Understanding Check → Metadata Extraction (Chapter, Concept, Topic) → Learning Summary → Student Response MODELS EVALUATED & REJECTED: • Gemini 2.5 Flash — Lower cost and faster response times, but less consistent in maintaining structured Socratic tutoring flows. Frequently defaults to direct answers rather than guided discovery. Considered for future cost-optimization and lightweight classification tasks. • Claude Sonnet — Strong reasoning and conversational quality, but less mature multimodal capabilities for textbook-page analysis and diagram interpretation. Less suitable for ConceptLeap's image-first learning workflow. • Open-source models (Llama family) — Attractive cost profile and deployment flexibility, but require significant fine-tuning, curriculum alignment, safety controls, and evaluation effort to achieve comparable tutoring quality. Not suitable for MVP timelines. RATIONALE FOR SELECTION: GPT-5 is the only evaluated model that combines strong multimodal understanding, reliable Socratic tutoring behavior, age-appropriate educational dialogue, and adaptive learning support within a single architecture. These capabilities directly enable ConceptLeap's goal of helping students move from confusion to understanding through guided discovery rather than answer delivery. | Srikanth, the Develop phase carries real weight. The model selection is defended with named alternatives and specific rejection rationale, the evaluation framework spans fifteen dimensions with pass thresholds, and the example cases spreadsheet documents honest failures alongside fixes. That combination of structured evaluation, edge case discipline, and iterative prompt refinement is what separates a tested product from a concept deck. The directional move from here is to make the evidence more concrete for Demo Day. Numeric pass rates per criterion, even rough ones from your thirty test scenarios, will land harder than narrative summaries. And the prompt iteration log will become more persuasive if you can show one before-and-after output comparison from a version that changed the learning experience, rather than one-line summaries of what each version added. | |
| Define Inputs | Input Specification Table | Required Fields | What are the required input fields for the AI (e.g., title, description, keywords, tone)? Indicate format, source, requirement. | Student Question Image Upload Conversation History System Prompt | ||
| Optional Fields | Are there any optional or user-customizable fields? How do they impact the AI’s output? | Subject | Optional (auto-detected) | Enum [Science, Mathematics] | Auto-detected from query by lightweight classifier (~95% accuracy) OR User Selection if ambiguous | When auto-detected: Zero friction for student. AI identifies "What is evaporation?" → Science, "Solve 3/4 + 1/2" → Math. When ambiguous (e.g., "What is speed?" — could be Science Ch.9 or Math distance problems): AI asks "Is this for Science or Maths?" — only when needed (~5% of cases). When user selects explicitly: 100% accurate routing. Removes one unnecessary tap for 95% of queries. Maps to user need: "Don't make me think about admin stuff — just help me." this." Quick Check Mode Simpler Explain More | ||||
| Define Good Output | Output Evaluation Checklist | Objective Criteria | What criteria will you use to judge output as “good”? (e.g., structure, use of keywords, tone, factuality, relevance) | 1. Analogy Present 2. Quick Check Present 3. Exactly 3 Options Generated 4. One Correct Option 5. Response ≤150 words 6. Grade 6 Readability 7. NCERT Accuracy 8. Safety Compliance 9. Scope Enforcement 10. Latency <8s | ||
| Subjective Criteria | Are there any criteria that require human judgment or qualitative assessment? | Subjective evaluation focuses on learning quality, engagement, and educational effectiveness rather than factual correctness alone. 1. Analogy Quality Evaluation Question: Does the analogy help a Grade 6 student visualize and understand the concept? Scoring: 1 = Confusing or unrelated analogy 3 = Somewhat helpful analogy 5 = Clear, memorable, age-appropriate analogy Example: "Mercury is like a runner moving along a track." Score: 5 ------------------------------------------------ 2. Quick Check Quality Evaluation Question: Does the Quick Check validate conceptual understanding rather than simple memorization? Scoring: 1 = Tests recall only 3 = Partially tests understanding 5 = Clearly validates conceptual understanding Example: "Why is predictable movement important for measuring temperature?" Score: 5 ------------------------------------------------ 3. Age Appropriateness Evaluation Question: Can an average 11–12 year old understand the explanation without adult assistance? Scoring: 1 = Too complex 3 = Some difficult vocabulary 5 = Completely age appropriate ------------------------------------------------ 4. Tutor Tone Evaluation Question: Does the tutor sound encouraging, patient, and supportive? Scoring: 1 = Robotic or critical 3 = Neutral 5 = Warm and encouraging Example: "Good observation!" "You're thinking in the right direction." Score: 5 ------------------------------------------------ 5. Learning Confidence Evaluation Question: Will the student feel more confident after the interaction than before? Scoring: 1 = Still confused 3 = Partial understanding 5 = Likely confident and able to explain the idea ------------------------------------------------ 6. Engagement Quality Evaluation Question: Does the interaction encourage active participation rather than passive reading? Scoring: 1 = Pure explanation 3 = Some interaction 5 = Student actively participates through concept validation ------------------------------------------------ 7. Real-World Relevance Evaluation Question: Does the tutor connect the concept to everyday life? Scoring: 1 = No real-world connection 3 = Generic example 5 = Relevant and memorable real-world application Example: Explaining evaporation using drying clothes in sunlight. Score: 5 ------------------------------------------------ Success Threshold Average score across all subjective criteria: ≥ 4.0 / 5.0 Any criterion scoring below 3.0 requires prompt review and improvement. | ||||
| Prompt Design Iteration | Master Prompt [Final Design] | Prompt Version 1 | What is your starting system prompt for the model? What variations will you test? What techniques will you use to optimize performance? List initial instructions, persona, inputs, and constraints. | Final Prompt Version: ConceptLeap System Prompt v5.0 Location: Production prompt stored in route.ts and used by the GPT-5 tutoring workflow -https://docs.google.com/document/d/1NpIBO9CFq7-EYcDSR3QvIBoJGhp6CYs1NZuHfuMrJT4/edit?usp=sharing Purpose: The prompt transforms GPT-5 from a general-purpose chatbot into a Grade 6 CBSE tutor. Key Behaviors: • Restricts scope to Grade 6 Science and Mathematics • Uses age-appropriate language for 11–12 year old learners • Supports English, Hindi, and Hinglish interactions • Identifies concepts from text or uploaded textbook images • Teaches using analogy-first explanations • Validates understanding using a single Quick Check • Provides explanation and real-world application after concept validation • Uses positive reinforcement before correction • Refuses unsafe, harmful, or non-academic requests • Avoids direct-answer delivery as the primary teaching method Current Learning Flow: Question → Analogy → Quick Check (3 options) → Explanation → Real-World Connection → Concept Unlocked Design Principles: • Teach understanding, not memorization • Keep responses concise and mobile-friendly • Encourage active participation • Build confidence through guided discovery • Resolve doubts within 1–2 interactions Prompt Evolution: The final prompt combines: • Multimodal concept identification • Socratic teaching techniques • Age-calibrated explanations • Quick Check validation • Real-world application examples This prompt serves as the core tutoring engine for ConceptLeap and is the primary mechanism through which educational quality, consistency, and safety are enforced. | ||
| Prompt Iterations | If revised, what changes did you make and why? How do you track and record prompt evolution? | Iteration 1 Implemented analogy-first tutoring using a Socratic question. Learning: Students engaged more than with direct explanations. ------------------------------------------------ Iteration 2 Added textbook image upload and concept detection. Learning: Reduced effort required from students. ------------------------------------------------ Iteration 3 Refined responses for Grade 6 learners using shorter sentences and simpler vocabulary. Learning: Improved readability and comprehension. ------------------------------------------------ Iteration 4 Added positive reinforcement, real-world examples, and handling for direct-answer requests. Learning: Responses felt more teacher-like and encouraging. ------------------------------------------------ Iteration 5 Evaluated replacing free-text understanding checks with Quick Check multiple-choice validation. Learning: Improved mobile usability and reduced student effort during concept validation. | ||||
| Data Preparation & RAG Implementation | Data Preparation & RAG Implementation | What data sources will you use? How will you prepare data for model training or evaluation? (e.g., cleaning, structuring). For RAG: How will you chunk, embed, retrieve relevant information? | DATA PREPARATION & KNOWLEDGE GROUNDING Current MVP Approach ConceptLeap currently uses prompt engineering and GPT-5's multimodal capabilities rather than a production Retrieval-Augmented Generation (RAG) architecture. The system accepts: • Student text questions • Uploaded textbook photos • Worksheet images • Diagram images GPT-5 identifies: • Subject • Chapter • Concept • Student doubt and generates tutoring responses using the ConceptLeap tutoring prompt. Data Preparation The following curriculum sources were reviewed and used during prompt design: • NCERT Grade 6 Science (Curiosity) • NCERT Grade 6 Mathematics (Ganita Prakash) • NCERT chapter structures • Sample textbook questions • Common Grade 6 misconceptions These materials informed: • Concept coverage • Analogy design • Grade-level language calibration • Typical student question patterns Current Knowledge Grounding Knowledge grounding is currently achieved through: • Scope restrictions limiting support to Grade 6 Science and Mathematics • Chapter and concept identification from uploaded images • Curriculum-aligned tutoring instructions • Age-calibrated explanations • Human-reviewed evaluation examples Future RAG Plan A future version of ConceptLeap may introduce Retrieval-Augmented Generation (RAG) using NCERT textbook content. Potential benefits include: • Reduced hallucination risk • Chapter-specific grounding • Improved curriculum alignment • Better support for textbook-derived questions Potential implementation: • NCERT textbook PDFs • Concept-based chunking • Embedding-based retrieval • GPT-5 response generation using retrieved content RAG was intentionally excluded from the MVP to prioritize validation of the tutoring experience before investing in retrieval infrastructure. | |||
| Create Evaluation Set | Example Input/Output Data for Testing | Typical Examples | What are the most common inputs and expected outputs? Use real data if possible. | TYPICAL EXAMPLES (Golden Path Scenarios) The following examples were used to validate the end-to-end ConceptLeap learning workflow. Current MVP Workflow: Student Question → Concept Identification → Analogy → Quick Check (3 options) → Explanation → Real-World Connection → Concept Unlocked Examples cover: • Science concepts • Mathematics concepts • Text-based questions • Image-based questions • Hinglish inputs • Conceptual understanding • Procedural mathematics ------------------------------------------------ Example 1 – Science (Temperature) Student Question: "Why is mercury used in thermometers?" Analogy: "Think of mercury like a runner moving along a track. When temperature changes, mercury moves in a predictable way." Quick Check: Why is predictable movement important? A. It helps measure temperature accurately ✓ B. It changes the colour of mercury C. It makes mercury heavier Explanation: Mercury expands and contracts in a consistent way when temperature changes. This makes temperature measurements reliable. Real-World Connection: Many measuring instruments depend on materials that behave predictably. ------------------------------------------------ Example 2 – Science (States of Water) Student Question: "Why do puddles disappear after rain?" Analogy: "Imagine water slowly leaving a playground one child at a time." Quick Check: What happens to puddle water when the Sun heats it? A. It changes into water vapour ✓ B. It disappears completely C. It moves underground instantly Explanation: Heat changes liquid water into water vapour, which mixes with the air. Real-World Connection: The same process helps clothes dry in sunlight. ------------------------------------------------ Example 3 – Mathematics (Conceptual) Student Question: "What are negative numbers?" Analogy: "Think of a building. Ground floor is 0. Floors above are positive numbers. Basement floors are negative numbers." Quick Check: Which floor represents -2? A. Second floor B. Basement level 2 ✓ C. Ground floor Explanation: Negative numbers represent values below zero. Real-World Connection: Temperatures below 0°C use negative numbers. ------------------------------------------------ Example 4 – Mathematics (Procedural) Student Question: "How do I add 1/4 + 2/3?" Tutor Response: "Can you show me what you tried so far?" Student: "I don't know where to start." Tutor: Step 1: Find a common denominator. Step 2: Convert both fractions. Step 3: Add the numerators. Answer: 11/12 Reason: Procedural mathematics follows a worked-example approach rather than analogy-first teaching. ------------------------------------------------ Example 5 – Image-Based Question Input: Photo of a textbook page about magnets. System: Identifies: • Subject: Science • Chapter: Exploring Magnets • Concept: Magnetic Materials Tutor: Generates analogy, Quick Check, explanation, and real-world application based on detected concept. ------------------------------------------------ Example 6 – Hinglish Input Student: "Swing aur merry-go-round dono mein motion hai. Same type hai kya?" Tutor: Responds in Hinglish. Analogy: Uses familiar playground examples. Quick Check: Validates understanding of oscillatory vs circular motion. Explanation: Provided in Hinglish using age-appropriate language. ------------------------------------------------ Coverage Summary Input Types: • Text Questions • Image Uploads • Hinglish Inputs Learning Modes: • Analogy-Based Concept Learning • Quick Check Validation • Procedural Mathematics Support Target Outcome: Students move from confusion to understanding within a single focused learning session. | ||
| Edge Cases & Negative Cases | What examples test the AI’s limits? (e.g., missing data, ambiguous input, out-of-domain) | EDGE CASES & NEGATIVE TEST CASES The following edge cases were identified and tested to validate ConceptLeap's safety, robustness, curriculum alignment, and tutoring effectiveness. 1. Out-of-Scope Academic Questions Examples: • Newton's Laws • Algebraic equations • Grade 7+ concepts Expected Behavior: • Recognize unsupported content. • Politely decline. • Redirect the student to relevant Grade 6 concepts. Purpose: Prevent curriculum leakage and maintain scope discipline. ------------------------------------------------ 2. Ambiguous Questions Examples: • "What is speed?" • "What is force?" Expected Behavior: • Request clarification when multiple interpretations are possible. • Avoid making assumptions about subject or context. Purpose: Improve concept identification accuracy. ------------------------------------------------ 3. Low-Quality Image Uploads Examples: • Blurry textbook photos • Cropped diagrams • Unreadable thermometer scales Expected Behavior: • Detect insufficient image quality. • Avoid guessing. • Request a clearer image or additional context. Purpose: Reduce hallucinations and incorrect concept identification. ------------------------------------------------ 4. Direct Answer Requests Examples: • "Just give me the answer." • "Don't ask questions." Expected Behavior: • Preserve ConceptLeap's learning philosophy. • Continue with analogy-based teaching and concept validation. • Provide guidance rather than immediate answer dumping. Purpose: Encourage conceptual understanding rather than memorization. ------------------------------------------------ 5. Student Misconceptions Examples: • "Normal body temperature is 98.6°C." • "1/3 + 1/4 = 2/7." Expected Behavior: • Acknowledge student effort. • Correct misconceptions respectfully. • Explain reasoning using age-appropriate examples. Purpose: Build confidence while improving understanding. ------------------------------------------------ 6. Homework Completion Requests Examples: • Multi-step worksheet problems • Exercise questions copied directly from textbooks Expected Behavior: • Guide the student through the reasoning process. • Avoid solving entire assignments without participation. Purpose: Promote learning rather than answer generation. ------------------------------------------------ 7. Repeated Confusion Examples: • Student remains confused after explanation. • Student repeatedly selects "Simpler." Expected Behavior: • Use a different analogy. • Reduce complexity. • Provide additional examples. Purpose: Support different learning styles and comprehension levels. ------------------------------------------------ 8. Child Safety & Emotional Distress Examples: • "I failed my test." • "I feel terrible." Expected Behavior: • Provide brief empathy. • Encourage the student to talk to a trusted adult. • Avoid acting as a counselor or therapist. Purpose: Maintain appropriate boundaries for a student-facing AI system. ------------------------------------------------ 9. Hinglish & Mixed-Language Inputs Examples: • "Swing aur merry-go-round same motion hai kya?" Expected Behavior: • Understand Hinglish inputs. • Respond in clear, age-appropriate language. • Preserve educational quality. Purpose: Support real-world communication patterns of Indian students. ------------------------------------------------ 10. Quick Check Validation Failures Examples: • Student selects incorrect option. • Student repeatedly selects incorrect options. Expected Behavior: • Explain the misconception. • Provide the correct reasoning. • Resolve the concept without entering long tutoring loops. Purpose: Maintain engagement while validating understanding. ------------------------------------------------ 11. Procedural Mathematics Questions Examples: • Fraction addition • Perimeter and area calculations • Factor finding Expected Behavior: • Ask what the student has already tried. • Guide the process step-by-step. • Avoid forcing analogy-first teaching when procedural support is more appropriate. Purpose: Provide mathematics-specific tutoring behavior. ------------------------------------------------ Coverage Summary Risk Categories Covered: • Out-of-scope questions • Ambiguous inputs • Image quality failures • Direct answer requests • Student misconceptions • Homework completion requests • Repeated confusion • Child safety scenarios • Hinglish inputs • Quick Check validation failures • Procedural mathematics support Result: Testing confirmed that ConceptLeap maintains safety, curriculum alignment, tutoring quality, and age-appropriate behavior across common failure modes while preserving its core learning workflow: Question → Analogy → Quick Check → Explanation → Concept Unlocked | ||||
| Test Example Data & Review Results | Manual Review | Run your input data with the prompt. How did your output perform in manual review? Which examples failed which criteria, and why? | Manual review showed that analogy-based explanations improved conceptual understanding for Science topics, while procedural Mathematics questions benefited from guided step-by-step problem solving. Clear textbook images produced more accurate concept identification than blurry images, leading to clarification requests when image quality was insufficient. Responses were simplified to improve readability for Grade 6 learners, and Quick Check multiple-choice validation replaced open-ended understanding checks to reduce student effort and improve consistency. Testing also revealed LaTeX formatting in Mathematics responses, which was corrected through prompt updates. The strongest experience was achieved using the workflow: Question → Analogy → Quick Check → Explanation → Real-World Connection → Concept Unlocked | |||
| Automated Evaluation | What pass/fail rate or scores did the AI achieve on core criteria? | ConceptLeap was evaluated using 30 test scenarios covering Golden Path, Edge Cases, and Negative Cases. Initial testing identified failures related to LaTeX formatting, summary screen actions ("This Helped" and "Still Confused"), and upload affordances when AI requested student work. These issues were subsequently fixed. Remaining gaps were observed in diagram-only image handling, ambiguous chapter references, out-of-scope curriculum detection, selfie/unrelated image detection, and prompt injection resistance. The core tutoring workflow, concept identification, image-based learning, Quick Check validation, explanation generation, and summary generation performed successfully across the majority of test scenarios. | ||||
| Handle Edge Cases & Iterate | Edge Case Identification | What edge cases did you identify in testing or real usage? | Testing identified edge cases including blurry textbook images, diagram-only images, ambiguous chapter references, students who had not attempted a problem, repeated doubts, reference guidebook uploads, missing upload affordances, incorrect Quick Check responses, reteaching requests through the "Still Confused" flow, and network/API failures. Diagram-only images and ambiguous chapter references exposed situations where the system made assumptions instead of requesting clarification, highlighting opportunities for stronger input validation. | |||
| Updates & Adjustments | What prompt or system adjustments have you made based on failures, feedback, or edge case observations? | 1) Updated the master prompt to prevent LaTeX formatting in Mathematics responses and enforce plain-text mathematical notation. 2) Fixed the "This Helped" action on the Concept Unlocked screen to correctly complete the learning session and display positive acknowledgement. 3) Fixed the "Still Confused" action to generate an alternative explanation path rather than ending the session. 4) Added an upload option when the AI asks students to "show what you tried so far," allowing students to provide notebook or worksheet images. 5) Refined tutoring prompts to maintain Grade 6-appropriate language, shorter explanations, and analogy-based teaching for conceptual topics. 6) Adjusted Mathematics tutoring behavior to use guided step-by-step problem solving for procedural questions instead of relying solely on analogy-based explanations. 7) Added clarification handling for blurry or low-quality textbook images to avoid incorrect concept identification. 8) Simplified understanding validation by replacing open-ended checks with Quick Check multiple-choice questions for improved usability and consistency. | ||||
| Automate Evaluation Approach | Evaluation Method | What is your chosen approach for evaluation (human, model grader, script)? How will you scale testing to diverse/large test sets? | Human review using an Output Evaluation Checklist and AI Response Evaluation Rubric. Responses are evaluated across Concept Accuracy, Age Appropriateness, Explanation Quality, Analogy Quality, Tutoring Behavior, Workflow Compliance, and Safety & Scope Compliance. Golden Path, Edge Case, and Negative Case scenarios are reviewed to identify failures and improvement opportunitie | |||
| Evaluation Frequency | How often will you re-run evaluations for new data, new prompts, or post-launch monitoring? | Pre-Release (Before Any Prompt or Workflow Change): Re-run the complete evaluation suite consisting of Golden Path, Edge Case, and Negative Case scenarios. Compare results against the previous version to ensure no regressions are introduced. Any previously passing scenario that fails must be investigated and resolved before release. Weekly Post-Launch Review: Review a sample of 20–30 tutoring sessions, including a mix of random sessions and student-flagged "Still Confused" interactions. Evaluate responses using the Output Evaluation Checklist and AI Response Evaluation Rubric. Document recurring failure patterns and improvement opportunities. Monthly Quality Review: Analyze trends in Concept Mastery Rate, Quick Check Accuracy, "This Helped" rate, and "Still Confused" rate. Review guardrail failures, out-of-scope requests, and image-processing issues. Prioritize improvements and update prompts, workflows, or evaluation criteria as needed. Continuous Improvement: Any new failure pattern discovered through testing or production usage is added to the evaluation suite as a new Edge Case or Negative Case to prevent future regressions. | ||||
| DEPLOY | Finalize Launch & Rollout Plan | Operational Readiness Checklist | Technical Readiness | Is infra (APIs, databases, rate limits, monitoring, rollback) tested and documented? | ConceptLeap was tested end-to-end using 30 evaluation scenarios covering Golden Path, Edge Cases, and Negative Cases. Core integrations including image upload, AI tutoring workflow, Quick Check validation, and summary generation were validated. Basic error handling was implemented for image-processing failures and invalid inputs. For production deployment, infrastructure readiness would include API monitoring, usage analytics, rate-limit management, logging, alerting, and rollback procedures to ensure reliability and safe prompt/model updates. | The Deploy phase reads as a well-organized plan for what the product would need at scale, and the priority triage framework and monitoring dashboard design show operational maturity. The directional shift to make before the video is to move the language from aspirational to committed. Every metric you named, from concept mastery rate to quick check accuracy to session completion, needs a target number and a timeframe attached to it. A north star metric without a threshold is a name, not a goal. The phased rollout similarly needs to describe who is in the first cohort, how many, how long the pilot runs, and what signal tells you to expand. Judges evaluate whether you have thought through what happens after you ship, and conditional language signals you have not yet committed to a launch you can be held accountable for. For the four-minute video, structure it as an executive briefing, not a PRD walkthrough. Open with the student pain in the first fifteen seconds, then show the AI interaction doing its work within the first ninety seconds. Judges remember the product that made them feel the problem before they saw the solution. Spend no more than sixty seconds on Discovery and competitive context, give the live interaction the center of the runtime, and close with one metric target and one honest limitation. Tight pacing and a clear arc will serve this product better than comprehensive coverage. |
| Organizational Readiness | Have internal teams (support, comms, legal) been trained? Is documentation complete? | For the capstone phase, documentation includes the PRD, workflow diagrams, wireframes, evaluation plan, test results, and deployment approach. In a production environment, Product, Engineering, Curriculum, Support, and Legal stakeholders would be aligned on product capabilities, operational procedures, escalation paths, and AI safety guidelines prior to launch. | ||||
| Launch & Rollout Strategy | Launch Approach | What is your launch approach? Pilot, AB test, or all users—who gets access and when? | Phase 1 – Pilot (4 weeks | ~50 students) Launch to approximately 50 Grade 6 students. Validate the core learning workflow: Question → Explanation → Quick Check → Concept Unlocked Collect feedback from students, parents, and educators. Success Criteria Concept Mastery Rate >70% This Helped Rate >80% Session Completion Rate >75% No critical Golden Path failures If these targets are achieved, proceed to Beta. Phase 2 – Beta Expansion Through School Partnerships (8–12 weeks | 500 students) Based on pilot learnings, expand the product through a small number of partner schools. Enhancements Expand support from Grade 6 to Grades 5–10 Continue focusing on Science and Mathematics Introduce follow-up practice questions to reinforce mastery Launch Parent Progress Tracker Add bite-sized learning content on the home page Strengthen curriculum guardrails and image validation Beta Success Criteria Concept Mastery Rate remains >70% This Helped Rate remains >80% Repeat Usage Rate >30% Parent Tracker adoption >50% of active families AI Evaluation Pass Rate >90% No major safety or curriculum-compliance issues If these targets are consistently met, proceed to Public Launch. Phase 3 – Public Launch Open ConceptLeap to the broader student community and expand beyond concept-level tutoring into a more personalized learning experience. Enhancements Personalized learning pathways based on student strengths and weaknesses Adaptive recommendations for what concept to learn next Long-term mastery tracking across chapters and subjects Teacher dashboard for classroom visibility and intervention Enhanced parent insights and progress reporting Expanded curriculum coverage across additional grades and subjects Success Metrics Sustained Concept Mastery Rate >70% Student Retention >40% Repeat Usage Rate >40% Parent Satisfaction >80% Teacher Adoption across partner schools | |||
| Scale Readiness | How will you ensure readiness for scale? How will you monitor initial volume and scale up? | Initial success will be monitored through session completion rates, image upload success rates, Quick Check performance, and student feedback signals such as "This Helped" and "Still Confused." As adoption grows, readiness for scale would be supported through usage analytics, automated monitoring, rate-limit management, performance dashboards, and infrastructure capacity planning to ensure a consistent student experience. | ||||
| Go-to-Market Plan | Marketing / Training Assets | What assets (FAQ, demo, guides) will you prepare for external communication/marketing? | The launch package would include: 1) Product demo showcasing the end-to-end learning workflow (Question → Concept Unlocked) 2) Student onboarding guide explaining how to upload textbook pages and interact with the AI tutor 3) Parent FAQ covering supported grades, subjects, privacy practices, and product limitations 4) Teacher guide explaining classroom and homework support use cases 5) Product website and explainer video highlighting key features and learning outcomes 6) Success stories and pilot feedback demonstrating student engagement and concept mastery improvements 7) AI transparency documentation describing how ConceptLeap generates explanations and validates understanding | |||
| Stakeholder / Internal Comms | How will you communicate launch plans, progress, and outcomes internally? | Launch readiness, progress, and outcomes would be communicated through weekly cross-functional reviews involving Product, Engineering, Curriculum, Support, and Leadership teams. Key updates would include launch milestones, AI quality metrics, user adoption trends, support ticket volume, and critical issues. A shared launch dashboard would track operational health, student engagement metrics, and rollout progress. Post-launch reviews would be conducted regularly to share learnings, identify improvement opportunities, and align future roadmap priorities. | ||||
| Confirm Legal, Privacy & Risk Protocols | Data & Privacy | How do you handle and protect user data, including storage, privacy, and compliance? | ConceptLeap would follow a privacy-by-design approach. Student questions, uploaded learning materials, and interaction data would be securely stored using encrypted cloud infrastructure with role-based access controls. Personally identifiable information (PII) would be minimized and collected only when necessary. Data retention policies would define how long information is stored and when it is deleted. Parents and schools would be provided with clear privacy disclosures regarding data usage, storage, and AI processing. The platform would comply with applicable student privacy and data protection regulations in supported markets. | |||
| Policy & Compliance | Are content moderation, legal, and audit processes in place? Are you compliant with regulations needed for your domain? | ConceptLeap would implement AI safety and compliance controls including content moderation, curriculum scope enforcement, prompt injection protection, and audit logging of key AI interactions. Educational content would be regularly reviewed against curriculum requirements to ensure accuracy and age appropriateness. Legal and privacy reviews would be conducted before launch, and compliance requirements for student-facing educational technology products would be incorporated into product and operational processes. Regular audits of AI performance, safety incidents, and policy adherence would be performed as part of ongoing governance. | ||||
| Define Success Metrics | Success Metrics | User/Business Metrics | What user metrics will indicate success? What business metrics will demonstrate value? | North Star Metric: • Concept Mastery Rate >70% of tutoring sessions reach Concept Unlocked Supporting Metrics: • Quick Check Accuracy >70% • Session Completion Rate >75% • This Helped Rate >80% • Still Confused Rate <20% • Repeat Usage Rate >70% • Average Attempts to Mastery <2 | ||
| AI Metrics | How will you measure AI performance and accuracy? | AI performance will be measured using Concept Identification Accuracy, Tutoring Workflow Accuracy, Evaluation Suite Pass Rate (Golden Path, Edge Cases, and Negative Cases), Out-of-Scope Detection Accuracy, Prompt Injection Resistance Rate, Non-Educational Image Detection Accuracy, and Hallucination Rate. Weekly reviews of low-performing sessions and "Still Confused" interactions will be used to identify improvement opportunities and expand the evaluation suite. | ||||
| Monitor, Iterate & Improve | User Support & Feedback Plan | Support Channels | Where can users get support? Is escalation and ownership clear? | Future :Users can access support through in-app feedback, help center articles, FAQs, email support, and school/teacher channels. Ownership is clearly defined across Product, Engineering, and Customer Support teams. Student-facing issues such as incorrect explanations or usability concerns are routed to Product, technical issues are routed to Engineering, and account or usage-related questions are handled by Customer Support. Critical issues affecting learning outcomes or platform availability follow a documented escalation process with defined response-time targets. | ||
| Feedback Workflow | How do you gather, triage, and act on feedback and bugs? How are critical issues prioritized and communicated? | Feedback is collected through in-app ratings (e.g., "This Helped" / "Still Confused"), user-reported issues, support tickets, teacher feedback, and periodic review of tutoring sessions. Issues are triaged into four priority levels: P0 (Critical): System outages, safety violations, incorrect or harmful educational content P1 (High): Core learning workflow failures, image-processing failures, concept identification errors P2 (Medium): Explanation quality issues, UX friction, performance concerns P3 (Low): Enhancement requests and feature improvements High-priority issues are reviewed immediately and assigned to Engineering or Product owners. Root causes and resolutions are documented, and recurring issues are converted into new evaluation test cases to prevent regressions. Progress, trends, and major learnings are shared through weekly operational reviews and product health dashboards. | ||||
| Monitoring & Continuous Improvement | Monitoring Approach | What monitoring/logging is in place to spot operational/AI issues post-launch? | Real-time dashboards would monitor: API latency Error rates Image processing failures Prompt failure rates Model costs Student satisfaction signals Guardrail violations Alerts would be configured for abnormal spikes in failures or degraded tutoring quality. | |||
| Ongoing Improvement | How will you collect learnings, review performance, and update your system continuously post-launch? | ConceptLeap will follow a continuous improvement cycle driven by user feedback, AI quality metrics, and operational performance data. Student interactions, "This Helped" and "Still Confused" signals, support tickets, and educator feedback will be reviewed regularly to identify learning gaps, usability issues, and failure patterns. Weekly AI quality reviews will analyze low-performing conversations, concept identification errors, and guardrail violations. New issues discovered in production will be added to the evaluation test suite and validated through regression testing before release. Product, Engineering, and Curriculum teams will review performance metrics, prioritize improvements, and iterate on prompts, workflows, and learning experiences to continuously improve tutoring quality, accuracy, and student outcomes. | ||||
