Weaver

• 

  logo

Inspiration

I wanted to write freely without needing organizing every detail afterward.

What it does

Weaver extracts people, facts, plans, and tasks from journal entries and quick notes. It builds people profiles, separates personal tasks from obligations involving others, and gathers reminders in Insights.

Users can manage dates, repeats, and task status. Clear completion statements can resolve matching one-off personal tasks. Extracted details retain exact source quotes.

How we built it

Weaver uses React, TypeScript, FastAPI, and SQLite. NVIDIA Nemotron 3 Nano handles extraction and resolution through Nebius Token Factory.

Token Factory provided hosted inference without requiring a GPU server. Local code validates results and handles scheduling, search, and completion matching. The demo uses isolated temporary sessions.

Challenges we ran into

Long entries exposed missed actions, ambiguous pronouns, duplicate reminders, and incorrect actor attribution. I addressed these through source validation, clearer prompts, targeted retries, and clause-level completion checks.

Accomplishments that we're proud of

Weaver connects natural writing to profiles and actionable reminders while keeping the journal central. The backend has 422 passing tests, alongside frontend tests and live-model checks.

What we learned

Hosted inference accelerated development, but reliable behavior required application-level checks. Nemotron needed safeguards for omissions, timing, and attribution.

What's next for Weaver

Improve completion matching for person-related promises and recurring tasks, reduce processing time, strengthen extraction coverage, and add optional persistent accounts.

Built With

• fastapi
• nemotron3
• react
• sqlite
• typescript