Learn‑KuchBhi

Inspiration

We learn best when concepts are explained visually and reinforced through practice. We wanted a way to turn any topic into bite‑sized study assets like animated explanations, organized flashcards, and a quick multiplayer quiz, so that students can go from curiosity to understanding in minutes.

What it does

AI‑powered Manim videos: Type a prompt, get a clean, educational animation. The system upgrades prompts, generates Manim code, lint‑fixes it, and renders an MP4 you can stream instantly.
Smart flashcards with outlines: It first designs a topic/subtopic outline, then generates high‑quality flashcard sets for each subtopic. Progress streams live via Server‑Sent Events (SSE).
Real‑time quiz (two‑player or solo): Play a fast multiple‑choice quiz on an AI‑generated topic set or math mode. WebSockets power low‑latency gameplay and score updates.
Accounts and persistence: Authenticated users can trigger generations; all runs, videos, and flashcards are stored and queryable.
Hosted demo: Live at https://kuch-bhi.built.systems

How we built it

Backend: FastAPI app with modular services for video generation, flashcards, and quiz. Uvicorn runs the ASGI server.
Video pipeline: Prompt upgrader (Gemini) → Manim code generator (Gemini/OpenRouter via pydantic‑ai) → Ruff lint/auto‑fix → preflight → render with Manim/MoviePy → store and serve under /videos.
Flashcards: Outline agent + per‑subtopic flashcard agents (pydantic‑ai). Stores runs and sets in PostgreSQL, streams run status and new sets via SSE.
Quiz: In‑memory manager with REST + WebSocket endpoints; supports topic‑AI or math modes, two players max, auto‑start when all players are ready.
Data & auth: SQLAlchemy models + Alembic migrations on PostgreSQL; FastAPI‑Users for auth with JWT (jwcrypto/pyjwt). Redis for sessions/caching.
Ops: uv for Python envs, Docker Compose for local Postgres/Redis, Nginx for TLS, SSE and WS proxying. Optional Context7 integration to fetch docs for the codegen agent.

Challenges we ran into

Making LLM‑generated Manim code reliably lint‑clean and runnable; added iterative lint/runtime fix loops and fallbacks.
Containing dependencies per‑session for rendering while keeping runs fast and deterministic.
Designing an outline‑first flashcard flow that scales across many subtopics with live progress via SSE.
Building a simple but robust WebSocket quiz loop with fair scoring, timeouts, and auto‑start while handling disconnects.
Schema evolution and status tracking for long‑running jobs across DB tables.

Accomplishments that we're proud of

A full end‑to‑end video pipeline that self‑corrects (lint/runtime) and produces clean animations from plain prompts.
Multi‑agent flashcard generation with persistent runs, per‑subtopic sets, and real‑time progress streaming.
A snappy two‑player quiz with both AI‑topic and math modes over WebSockets.
Production‑ready touches: JWT auth, DB migrations, static video serving, Nginx WS/SSE proxying, and environment‑based provider selection.

What we learned

How to steer LLMs for deterministic, lint‑clean code generation with feedback loops.
Practical patterns for SSE and WebSockets on FastAPI behind Nginx.
Schema design for content generation pipelines and incremental result surfacing.
The trade‑offs between provider models (Gemini vs OpenRouter) for different tasks.

What's next for Learn‑KuchBhi

Rich frontend for creation, review, and study flows; shareable links and classrooms.
Editable timelines for videos and flashcard curation; export to Anki/CSV.
More quiz modes (buzz‑in, timed rounds), leaderboards, and persistence.
Caching, deduplication, and reuse of assets; cost/latency optimizations.
More providers and guardrails; better observability and retry logic.