In production
Orbit
Upload notes and Orbit builds flashcards and adaptive quizzes from them, grounding every question in your own material and focusing on what you keep getting wrong.
- Role
- Full-stack & AI engineer
- Timeline
- [n] months
- Year
- [2024]
- Stack
- React, Python, LangChain
(01)The problem
Students spend more time making study material than studying it.
Generic AI quizzes drift away from the syllabus — learners need questions grounded in their own notes, not the internet's.
(02)My role
- Built the ingestion pipeline: parsing, semantic chunking, embedding and indexing in Pinecone.
- Designed generation prompts and a validation step so every card cites its source chunk.
- Implemented spaced repetition with adaptive difficulty driven by answer history.
(03)Architecture
Notes are chunked, embedded and stored per user in Pinecone namespaces. Card and quiz generation is retrieval-augmented: the model only sees retrieved chunks, and each output is checked against its source before it is saved.
- 01
Ingest
PDF / Markdown / text → semantic chunks
- 02
Embed & index
Per-user Pinecone namespaces
- 03
Generate
RAG prompts → cards & quiz items
- 04
Verify
Grounding check against source chunk
- 05
Adapt
Spaced repetition + difficulty model
Grounding check before save
A second, cheaper model call verifies each card against its chunk — trading a little cost for much lower hallucination rates.
Namespaces per user
Isolates data cleanly and keeps retrieval fast without a metadata filter on every query.
(04)Tech
AI
- LangChain
- OpenAI
- Pinecone
- Embeddings
Backend
- Python
- FastAPI
- PostgreSQL
Frontend
- React
- TypeScript
(05)Outcome
[Learner outcomes, usage, or a quote from a user.]
- cards from a page of notes
- 10×
- grounded-answer rate
- 92%
- faster revision
- 3×
Next case study
Distributed Workflow Orchestrator
Self-hosted durable execution, Temporal / Inngest-style