# Studywise — product thesis and industry-readiness plan

**Updated:** 13 October 2026  
**Stage:** Hackathon prototype moving toward a pilotable product  
**Working product name:** Studywise (name/trademark clearance has not been done)

## 1. The market reality

The basic category is established. NotebookLM offers source-grounded study guides, flashcards, quizzes, citations, and audio/video learning artifacts. Quizlet converts teacher notes, slides, and readings into practice tests, study guides, and flashcards. StudyFetch markets a broad suite built around uploaded material, practice tests, and a conversational tutor. Khan Academy continues to invest in guided tutoring and interactive diagrams. A new product cannot credibly claim differentiation merely because it accepts files, cites sources, or generates a quiz.

| Product | Publicly described strengths | Implication for Studywise |
|---|---|---|
| NotebookLM | Source-centered synthesis, citations, study artifacts, quizzes/flashcards, audio/video overviews | Source grounding and convenient artifact generation are table stakes. |
| Quizlet | Turns notes, slides, and readings into practice tests and study guides; class sharing | Practice content is a mature and familiar workflow. |
| StudyFetch | Multiple study tools from one source set, practice-test maker, conversational tutor | A wide tool menu alone is easy to copy and hard to defend. |
| Khanmigo / Khan Academy | Guided tutoring, course content, and continued investment in interactive diagrams | Pedagogy and useful guided help matter more than a chat box. |

These are high-level observations from public product pages, not a feature-by-feature teardown, pricing study, or independent quality benchmark. The landscape moves quickly; re-check it before a public launch.

## 2. Positioning to test

> **Studywise helps a student find the exact course idea they misunderstood, repair that misconception from the course material, and prove the idea transferred to a new question.**

The product wedge is a **course evidence-to-mastery loop**:

1. Preserve each extracted concept with its page, slide, or video timestamp and a link to the original.
2. Track more than a single mastery percentage: keep evidence for each estimate (answered items, confidence, source, recency, and misconception hypotheses).
3. Diagnose a likely misconception from a wrong answer, show the student the relevant source excerpt, and teach the distinction with a short, source-grounded repair lesson.
4. Ask a new transfer question, not a reworded repeat, and update the concept estimate from the result.
5. Let students and instructors inspect why a recommendation was made and correct bad topic tags or source mappings.

Competitors already offer overlapping individual capabilities. **The defensible hypothesis is the quality and auditability of this connected workflow**, not that every component is unique. Validate the workflow with students and educators before making exclusivity claims.

### Initial audience and beachhead

Start with university students studying one course from a mix of instructor slides, lecture recordings, and textbook readings, initially in high-enrollment STEM courses. They have a clear study cycle, concrete source locations, and assessable concepts. Avoid launching simultaneously into K–12 (guardian, safeguarding, and district requirements), every subject, and institution-wide LMS procurement.

### Product principles

- Evidence before eloquence: show the cited passage, source location, and any missing evidence.
- Teach toward retrieval and transfer: explanations should end with a learner action, not only more generated content.
- Make the learner model inspectable and revisable; never present an estimate as an objective grade.
- Student owns their learning record. Export and delete must be visible and dependable.
- Separate a course source from instructions embedded inside that source. Uploaded material is untrusted data.
- Never use fabricated learning gains, customer logos, endorsements, or benchmark scores in marketing.

## 2a. Differentiation bets to explore

| Bet | Student value | Why it could stand out | Main risk / validation |
|---|---|---|---|
| **Evidence receipt for every mastery change** *(prototype now)* | “Why did my estimate move?” has a clear answer, outcome, timestamp, and reopenable source | Makes personalization inspectable instead of a black-box score | Check whether learners use it and whether citations remain correct after source updates. |
| **Misconception fingerprint, not just topic score** | Distinguishes errors such as confusing cause with mechanism, reversing a sequence, or mixing a definition with an example | Leads to a more useful repair lesson than “practice this topic” | Require educator-annotated error labels; unverified diagnosis could mislead. |
| **Course disagreement detector** | Flags when lecture, slides, and textbook appear to disagree, and presents both exact passages | Helps students navigate real course material rather than silently flattening conflicts | Detecting a true contradiction is hard; show uncertainty and ask the student/instructor to resolve it. |
| **Transfer ladder** | Separates remember, explain, apply, and connect questions for each concept | Measures whether the learner can use a concept in a new context, not only recognize a word | Needs validated item generation and human review; do not infer “mastery” from one cloze item. |
| **Portable course graph** | Exports concept links, sources, and review history in a documented format | Student-owned learning history reduces lock-in and can move between semesters/tools | Privacy-by-default export and source copyright controls need design. |

Prioritize the evidence receipt (already prototyped) and one instructor-reviewed misconception category in the pilot. Keep conflict resolution and deeper transfer assessment as research bets until their accuracy is measurable.

## 3. What is built now

The prototype includes a separate marketing page and a working course app, a seeded cell-biology demo course, browser-side source extraction for selectable-text PDFs, PPTX slide text, VTT/SRT transcripts and notes, lexical retrieval, cited passages, unsupported-question refusal, source-derived practice, a Bayesian mastery estimate, an evidence-linked spaced-review queue, exam-date/time-budget planning, a topic flow map, and a transparent small benchmark. It demonstrates the product hypothesis; it is **not production-ready**. See `README.md` for the current scope and run instructions.

## 4. Build plan and exit criteria

### Phase A — product story and prototype (current)

- Give the product a differentiated, accurate public narrative and an obvious route into the experience.
- Keep the demo complete with realistic course content and visible grounding.
- Document scope and limits honestly.
- Exit when an unfamiliar student can understand the promise and reach a useful action from the landing page without assistance.

### Phase B — validated student workflow

- Interview 8–12 university students and 3–5 instructors; observe real exam preparation, source switching, and wrong-answer review.
- Test three alternatives: “course-grounded tutor,” “fix what you misunderstood,” and “one learning map across lectures, slides, and readings.”
- Run a 2-week concierge pilot in one course with permission from its instructor. Use public/open or explicitly approved materials.
- Measure activation (first source indexed and first citation opened), first-session completion, weekly return, question usefulness, and student-reported trust. Do not infer learning from engagement.
- Pre-register a small learning study with an instructor: aligned pre/post assessment, comparison condition where feasible, adequate sample-size plan, and analysis of item difficulty. Do not publish causal claims from a convenience pilot.
- Exit when users repeatedly complete the misconception repair loop and can explain why it helped; revise the wedge otherwise.

### Phase C — learning quality and benchmark

- Replace keyword-only tagging with an editable course concept graph. Keep extraction confidence, provenance, prerequisites, and teacher corrections.
- Use a validated item schema: learning objective, stem, answer, distractors, rationale, difficulty, source span, source-derived answer check, and novelty ID.
- Add question-type-specific verifiers: symbolic/numeric execution with units and tolerances; rubric-based short-answer review with cited evidence and calibrated partial credit; distractor and ambiguity checks.
- Build a representative evaluation set with answerable, ambiguous, conflicting-source, and off-material queries. Annotate gold source spans, acceptable answers, expected refusal, and misconception labels.
- Run RAGAS/DeepEval or a comparable framework alongside blind educator review. Report per-category sample size, confidence intervals where useful, failure examples, model/version, and retrieval settings. Do not treat one aggregate score as proof of safety.
- Exit only when unsupported-answer rate, citation span accuracy, answer-key accuracy, and question repetition meet thresholds agreed with pilot instructors.

The review queue uses retrieval before reveal because controlled research found retrieval practice improved delayed recall relative to repeated study in the tested material. Spacing research shows that the useful gap depends on the desired retention interval; there is no single interval ladder that should be advertised as universally optimal. The current Again/Hard/Good/Easy schedule is therefore an inspectable starting heuristic. A pilot should compare schedules and use actual item-level history before making stronger claims.

### Phase D — production architecture

| Layer | Proposed responsibility | Minimum production decisions |
|---|---|---|
| Web app | React/TypeScript or equivalent accessible client, course workspace, learner controls | WCAG 2.2 AA target; keyboard, focus, responsive/reflow, captions, screen-reader review. |
| API | Authenticated course/source/session endpoints | Tenant and course authorization on every operation; schema validation, rate limits, audit events, idempotency. |
| Ingestion | Async workers for PDF, PPTX, OCR, image/diagram extraction, audio/video transcription | Malware scan, MIME/signature validation, size/page/duration limits, timeouts, parser isolation, status/retry/dead-letter. |
| Evidence store | Object storage for originals; Postgres for metadata and learner events; vector + lexical retrieval | Encryption in transit/at rest, per-tenant ACL filters, retention and deletion, backup/restore, source-version tracking. |
| Tutor service | Retrieval, evidence selection, constrained generation, structured citations, refusal checks | Prompt-injection isolation, no tool permissions by default, claim-to-span validation, model/provider abstraction and cost caps. |
| Learner service | Concept estimates and evidence ledger | Version the update algorithm; support uncertainty, confidence, recency, reset/correction, export, and explainability. |
| Operations | Metrics, tracing, moderation/escalation, incident response | PII-minimized logs, secrets management, dependency inventory, SLOs, alerts, disaster recovery, data residency review. |

Suggested first deployment: one region, one isolated student workspace per account, Postgres plus object storage, a managed background queue, and a model-provider adapter. Choose specific vendors only after data residency, educational-data requirements, unit economics, and a real pilot are known. No student upload should become model training data by default.

### Phase E — trust, privacy, and safety gates

- Threat-model direct/indirect prompt injection, cross-course data leakage, malicious files, parser exploits, unsafe generated links, resource exhaustion, and account takeover.
- Enforce per-course retrieval filters before generation. Cite only authorized source IDs. Never let retrieved document text change system policy or invoke tools.
- Put upload limits, file-type signatures, malware scanning, sandboxed parsers, OCR/time quotas, and safe error handling in place.
- Add clear consent, retention defaults, export/delete controls, deletion verification, incident contact, and a privacy notice reviewed for launch jurisdictions.
- Decide age policy and whether minors are in scope. If they are, design safeguarding, guardian/educator roles, data minimization, and applicable legal review before launch.
- Have domain educators review explanations and assessment items; define a process for reporting and correcting a bad source, citation, or answer key.
- Target WCAG 2.2 AA through automated checks plus manual keyboard and assistive-technology review; document remaining exceptions.

### Phase F — go-to-market and launch

- Launch a closed course pilot before self-serve scale. Publish the data policy, supported formats, limitations, and evaluation method.
- Build the demo around one compelling sequence: upload lecture + slides → open exact source citation → reveal a misconception → repair lesson → transfer question → explain mastery update.
- Recruit through instructors, course societies, and student study groups. Do not lead with generic “AI tutor” ads.
- Track activation and learning quality separately. Maintain a rollback switch for new extraction/model versions.
- Before public launch, clear product name and domain, contracts, privacy/security review, incident response, accessibility, support, and unit economics.

## 5. North-star and guardrail metrics

**North-star candidate:** weekly active students completing a *verified learning loop* (attempt → cited feedback → novel transfer question) on at least one concept.

Pair it with guardrails:

- Gold-span citation accuracy and source-opening success.
- Unsupported-query refusal precision/recall and contradictory-source handling.
- Independently reviewed item correctness, ambiguity rate, and duplicate rate.
- Concept-level pre/post transfer gain from a study design, not self-reported confidence alone.
- Time-to-first-useful-citation, upload/index completion, and source parse coverage.
- Weekly retention, help-seeking completion, and accessible completion rates by device and assistive technology.
- Privacy deletion completion, unauthorized retrieval incidents (target zero), and cost per verified learning loop.

## 6. Competitive research references

- Google, NotebookLM student learning features (flashcards/quizzes with citations, study guides, audio formats): [official product blog](https://blog.google/innovation-and-ai/models-and-research/google-labs/notebooklm-student-features/)
- Google, NotebookLM audio/video and source grounding: [official product blog](https://blog.google/innovation-and-ai/products/notebooklm-audio-video-sources/)
- Quizlet, AI-powered study tools: [official product page](https://quizlet.com/features/ai-study-tools)
- StudyFetch, AI study tools: [official product page](https://www.studyfetch.com/tools)
- Khan Academy, Khanmigo and interactive diagram developments: [Khan Academy announcement](https://blog.khanacademy.org/new-ai-tools-bring-interactive-diagrams-and-targeted-practice-thanks-to-khan-academys-partnership-with-google-org/)
- OWASP, LLM risks and mitigations including prompt injection and data access controls: [OWASP Top 10 for LLM Applications](https://genai.owasp.org/download/43299/?tmstv=1731900559)
- W3C, WCAG 2.2 guidance: [Understanding WCAG 2.2](https://www.w3.org/WAI/WCAG22/Understanding/)
- Karpicke & Roediger, retrieval practice and delayed recall: [Science (2008)](https://doi.org/10.1126/science.1152408)
- Cepeda et al., spacing and retention interval: [Psychological Science (2008)](https://www.psychologicalscience.org/journals/psychological-science/j.1467-9280.2008.02209.x)
- U.S. What Works Clearinghouse, study organization recommendations: [Organizing Instruction and Study to Improve Student Learning](https://ies.ed.gov/ncee/WWC/PracticeGuide/1)