ParikshaMind

50M+ Indians grind government exams — UPSC, SSC, state boards like Gujarat’s GSSSB — on textbooks and YouTube. Every AI tutor forgets them the moment the chat ends: it re-asks material they’ve mastered, has no memory of their weak spots, and no sense of progress. The thesis of ParikshaMind is that memory across sessions is the missing differentiator, not a bigger model.

Built solo inside the WeMakeDevs hackathon’s open-source, self-hosted Cognee track.

A FastAPI backend imports Cognee directly and owns every memory call against a fully self-hosted store — SQLite for relational, LanceDB for vectors, and a Ladybug graph, all local and file-based. Each of Cognee’s branded APIs does visible work: remember() ingests the syllabus and each quiz result into a per-learner dataset; recall() pulls a concise mastery summary at session start; search() with graph completion generates syllabus-grounded MCQs; improve() re-weights the learner’s graph after a session; forget() wipes it on request.

Per-learner isolation is just a user_<id> dataset, so re-entering the same name across a refresh recalls prior performance. Embeddings run locally via fastembed — no key, no cost.

The build started on Gemini 2.5 Flash, but the free key capped at ~20 requests a day — unusable for a live demo. I switched the reasoning LLM to Groq’s llama-3.3-70b-versatile free tier via litellm, rate-limited in config.

The real fix was structural: quizzes are pre-generated once, offline, into a question bank and served instantly at runtime, so the only tokens spent live are the light recall() call. The tutor feels instant and costs almost nothing to run.

• Cognee over rolling my own memory. A hybrid graph + vector store keyed per learner is exactly the shape this problem needs; reimplementing it would have burned the whole week.
• Pre-generated bank over live generation. Instant quizzes and near-zero runtime cost, at the price of variety per topic — a good trade for a memory demo, not for production.
• Groq over Gemini. Forced by a rate cap mid-build. Worth stating plainly: the shipped tutor runs on Groq, even though older planning docs still say Gemini.

A “mastery map” graph visualization was planned for the last day and didn’t ship —/progressreturns a text summary and a mastered-topics list, not a rendered graph. That’s the first thing I’d finish. After that: move the per-process session store into Cognee itself so runtime state survives a restart, and widen the question bank so mastery is earned across more variety, not the same eight questions per topic.