AI Learning Assistant — Text-to-Video Tutor

The problem

Two pain points sit on either side of "I want to learn this":

The reader's side. A textbook chapter or a stack of PDFs is fine for a determined student, but most people don't know what to learn first. There's no learning path baked into the material.
The teacher's side. Good teachers spend hours turning notes into a lecture video — scripting narration, finding diagrams, timing pacing, cutting captions. The step from "I know what I want to teach" to "I have a video of it" is where the energy runs out.

The idea

Treat written notes — or even a plain-text prompt — as a source, not a script. Build two cooperating systems:

A learning-path builder that reads documents, extracts knowledge points without duplicates, sequences them, and outputs a personalised plan with milestones.
A video tutor that takes any one of those knowledge points and turns it into a fully narrated, illustrated, bilingual lecture video.

What it does

Document → personalised course. Upload PDFs or DOCX. The system chunks them, embeds them into FAISS, extracts knowledge points, deduplicates them, sequences them into a learning path with milestones and assessments, and stores everything for later use.
Knowledge point → on-demand content. Click a KP; the system runs semantic search against the FAISS store for that source and a "Senior Teacher" persona generates explanation, examples, and practice material grounded in the original document.
Knowledge point → tutorial video. A multi-agent pipeline turns the KP into a narrated, illustrated Manim video — bilingual (EN/中文), with synced captions and exportable SRT/MP4/VTT.
KYC-driven personalisation. A LangGraph-based KYC pipeline tracks learner state across sessions: known concepts, knowledge gaps, learning style, available time.

How it's built

Pure LangChain backend. Document loaders (PyMuPDF, Docx2txt) → RecursiveCharacterTextSplitter (2000 / 100) → FAISS vector store with OpenAI embeddings → MMR semantic search.
Hybrid Contextual Batching for KP extraction. Small documents go through a single holistic LLM call; large ones go through context-aware batches where each batch is prompted with the previous batch's extracted KPs so duplicates fail at extraction time, not after.
Multi-agent video generation. Separate agents own Content (script + structure), Visual (Manim scenes, diagrams), and Error Handling (Manim syntax validation, retries). An MCP-style server ecosystem gives them one shared tool surface for rendering, audio synthesis, storage, and curriculum tooling.
AsyncIO streaming pipeline. Slide generation, audio synthesis, and scene rendering run in parallel through a dynamic worker pool — what used to take ~10 minutes drops to under 2.
Multi-model orchestration. GPT-5 / GPT-4 / Claude / Whisper, each routed where they're strongest, with fallback routing, exponential backoff, and AI-driven syntax correction for Manim errors.
FastAPI service. Endpoints for single-file processing, multi-file ingestion, KP content generation, learning-path retrieval, and processing status.

What I learned

De-duplicate at extraction time, not after. Cleaning up duplicate KPs post-hoc is fighting against the wrong layer; feeding context into the prompt is what actually fixes it.
The hardest part is pacing. A video that generates "perfectly" at double speed is unwatchable. Pacing is a first-class design concern, not an output setting.
Teachers care about structure, not voice. They tolerate a synthetic voice — they don't tolerate missing pauses, missing callouts, or "let's come back to this later" never coming back.
Export real formats from day one. SRT, MP4, VTT, JSON learning paths — exporting to real production formats turns a demo into a tool.

Where it goes next

Pause-to-quiz moments embedded inside the video, generated from the same source notes. A KYC-aware "next lesson" recommender that uses the Track Record of completed milestones to choose what to show next.