MapBrain
6-month freelance mission for an AI ed-tech startup: I built the engine that turns any content (PDF, YouTube, audio, slides) into a complete learning path — summaries, mind maps, quizzes, flashcards, podcast — plus a real-time voice tutor you can actually talk to about your course.
Private demo
This freelance project is confidential. The screenshots above are shared with client approval.
// recruiter view
Freelance mission (6 months) for MapBrain, an AI ed-tech startup: end-to-end design and delivery of an AI educational SaaS that turns any content (PDF, YouTube, audio, slides) into a personalized interactive learning path — multi-format generation, real-time voice tutor and full-Azure architecture on AKS — in close collaboration with the product team.
- ▸Freelance client mission delivered end-to-end: AI engine and backend owned solo, in close collaboration with the in-house product team and frontend
- ▸Multi-format pedagogical generation from one source document: summaries, mind maps, quizzes, flashcards and audio podcasts via LangGraph multi-agent workflows
- ▸Real-time conversational voice tutor — learners talk to their course through a Whisper → LLM → TTS pipeline streamed over WebSocket
- ▸Multi-modal ingestion (PDF, YouTube, audio, slides) with contextual RAG on Milvus across the learner's full content library
- ▸Full-Azure production architecture: AKS, Azure OpenAI, Redis, GitHub Actions CI/CD, monitoring and autoscaling
- ▸Shipped to production within a 6-month engagement, iterating fast with the product team on pedagogical quality
The story behind
- Chapter 01
The mission
MapBrain, an AI ed-tech startup, brought me in as a freelance AI engineer to design and build the heart of their product: an engine that turns any course material into a complete learning experience. Clean split of roles — I own the AI engine, the backend and the cloud deployment; the in-house team owns product direction and the frontend. One goal: ship a production SaaS in six months.
- Chapter 02
The problem
Learners are buried in raw material — PDFs, lecture recordings, YouTube lectures, slide decks — with no fast path from 'I have the content' to 'I can actually study this.' The product bet: from a single source, automatically generate every artifact a student needs (summary, mind map, quiz, flashcards, podcast) and let them ask questions about it out loud.
- Chapter 03
Multi-format generation
The core engine ingests one document and fans out into five pedagogical formats through LangChain/LangGraph multi-agent workflows on Azure OpenAI (GPT-4o). Each format is a specialized agent path with its own prompts and output schema — structured summaries, interactive mind maps, self-assessment quizzes, revision flashcards, and an audio podcast generated end-to-end.
- Chapter 04
The voice tutor
The standout feature: a real-time voice tutor. The learner speaks, Whisper transcribes, the LLM answers with full course context, and TTS streams the reply back — all over a WebSocket (Socket.IO) connection tuned for low latency and bidirectional audio. It turns a static course into a conversation.
- Chapter 05
Ingestion & RAG
To ground every answer in the learner's own material, I built a multi-modal ingestion pipeline: PDF extraction and chunking, Whisper audio transcription, YouTube transcript retrieval, slide parsing — all vectorized and stored in Milvus. Contextual RAG then spans the user's entire content library, not just the document in front of them.
- Chapter 06
Full Azure & async jobs
Everything runs on Azure: containers on AKS, Azure OpenAI for the models, Redis for sessions and caching, GitHub Actions for CI/CD, with autoscaling and monitoring. Long-running tasks like podcast synthesis (2–5 minutes) moved to an async job API — submit, poll status, download — so the frontend never hangs on a timeout.
After going live
Monitoring & observability
Production observability on Azure: AKS metrics and autoscaling, GitHub Actions pipelines, and latency tracking across the generation and voice pipelines.
Stack
Tracked metrics
- ●Azure OpenAI latency per format (summary / mindmap / quiz / flashcards / podcast)
- ●Voice tutor round-trip latency (Whisper → LLM → TTS)
- ●Async job duration and success rate (podcast synthesis)
- ●Milvus retrieval latency and RAG hit quality
- ●AKS pod autoscaling and resource utilization
Production impact
Shipped to production within the 6-month freelance engagement: a working AI educational SaaS turning any content into a personalized, interactive learning path with a real-time voice tutor.
- ▸End-to-end AI engine and backend delivered for a client product, from multi-modal ingestion to a real-time voice tutor
- ▸Five pedagogical formats generated automatically from a single source document
- ▸Real-time conversational voice tutor in production over WebSocket
- ▸Sovereign full-Azure architecture (AKS + Azure OpenAI) built to scale
- ▸Fast iteration loop with the product team on pedagogical quality
// results
// stack