Thenali AI
Developer Learning & Repository Intelligence
Reading a large open-source repo cold is a wall. Docs lie, READMEs are stale, and the actual architecture only lives in the code. Thenali AI turns any repo into a paced, explainable learning path grounded in the real source tree.
- Role
- Backend + AI pipeline lead — ingestion, embeddings, roadmap generation.
- Timeline
- 10 weeks
- Source
- GitHub ↗
What was hard.
Making explanations faithful to the code
LLMs summarise confidently and wrongly. Every explanation is tied back to a file:line range retrieved from FAISS, so users can verify against source in one click.
Adaptive pacing without a user model
Cold-start learners have no signal. The roadmap starts with entrypoints + config, then branches by module coupling — a graph-first curriculum, not a chapter list.
Multilingual without translation drift
Explanations are generated in the target language directly via Bedrock rather than post-translated, preserving code term accuracy. The interface supports 12 Indian languages natively.
How it fits together.
Repo → chunk → embed → roadmap → chat
- ▸S3 holds raw repos; FAISS index is per-repo and rebuilt on push.
- ▸DynamoDB tracks learner progress cheaply at read-heavy scale.
What shipped.
- —Grounding beats fluency. Users trust an ugly citation more than a smooth paragraph.
- —Entrypoint-first roadmaps mirror how senior engineers actually read code.
What the evidence supports.
Evidence status
The file-and-line-grounded explanation approach, the repository-to-roadmap pipeline, and the 12-language interface UI are working features. First-lesson timing and citation coverage are listed as design targets; they have not been measured in a formal evaluation.