Forgetwell is a capture-first note app. People save short, incomplete fragments without filing them, then retrieve them by exact words or meaning. It is a second memory, not a system that organizes or thinks for the user.
This role focuses on retrieval quality over the notes people deliberately capture. Improve embeddings, ranking, multilingual recall, and query understanding so time, entities, and question-shaped searches return useful fragments instead of a list of vaguely similar notes. The data boundary matters: a connected assistant may search or write notes only after the user explicitly connects it and grants the corresponding permission.
What you will do
- Own exact and meaning-based retrieval quality end to end: models, chunking, ranking, and evaluation
- Make recall understand intent — time ranges, entities, and question-shaped queries — not just vector similarity
- Build evaluation sets from messy, fragmentary notes and measure retrieval quality before tuning it
- Improve multilingual search and meaning-based recall on real user vocabularies
- Evolve optional MCP tools so assistant read and write capabilities remain permission-bound and legible
- Keep inference fast and affordable at single-binary scale
- Protect capture speed and the calm product experience while retrieval grows more capable
What we look for
- You have shipped retrieval, search, or embedding systems to production
- You have real opinions about privacy-preserving ML: data minimization, user-controlled processing, and on-device or local inference where it fits
- You can evaluate retrieval over short, incomplete, real-world text
- You measure before you tune, and you distrust vibes-based quality claims
- You can work close to the product: latency budgets, UX, and privacy constraints
- Bonus: intent parsing and query understanding, multilingual NLP, on-device or small-model inference, MCP or agent tooling