deepwork.ai
Built: July 18, 2026
Most focus apps impose the same system on everyone: 25 minutes on, 5 minutes off, repeat. deepwork.ai starts from a different premise — your focus patterns are yours, and a coach should learn them from your actual behavior instead of prescribing a generic ritual.
So it logs real session data, and puts an LLM agent on top that's only allowed to talk about what the data actually shows.
The focus timer
A Pomodoro-style timer that records more than start/stop:
- Distractions — what pulled you out, and when
- Pauses and completion state for every session
- Location — Home / Office / Cafe / Other, captured via a Leaflet map + geocoding
All of it lands in Supabase tables (focus_sessions, session_events, distractions) that become the raw material the coach reasons over.
The coach that can't make things up
The AI coach never sees a raw database dump. It's a tool-calling agent (OpenRouter → Gemini 2.5 Flash) that has to go fetch grounded facts before it's allowed to say anything specific about you:
get_focus_trends,get_best_focus_windows,get_distraction_patternsget_focus_by_location,get_distractions_by_locationget_recent_changeslog_coach_insight— the one write tool, for saving what it learns
Each chat turn runs a tool loop of up to 10 iterations: the model picks tools, the server executes them against Supabase RPC functions, results feed back in, and only then does it answer. The system prompt explicitly forbids fabricating metrics — the coach can push you toward small, testable experiments, but it can't invent a claim about your data it hasn't actually queried for.
Three selectable personalities sit on top of the same grounded core: strict Alex, data-focused Morgan, encouraging Sam.
Proactive, not just reactive
Beyond the chat, an autonomous agent layer runs in the background — surfacing insights, suggesting sessions, and generating weekly reports without being asked. A daily Vercel cron job refreshes derived analytics tables (daily_focus_stats, weekly_focus_patterns, focus_anomalies, coach_memory) so the coach is always reasoning over up-to-date aggregates rather than recomputing from scratch on every request.
Built to be measured
Built for the Opik (Comet) hackathon, so every LLM call is traced end-to-end, backed by eval datasets and a before/after experiment comparing an early regression prompt against the tuned production one — the kind of evidence that's usually missing from "the AI got better" claims.
Stack
- Next.js 15 (App Router), React 18, TypeScript
- Supabase — Postgres, auth, RLS, RPC functions as agent tools
- OpenRouter → Gemini 2.5 Flash for the coach
- Opik — LLM tracing and evaluation
- react-leaflet — location capture and maps
- Vercel — hosting + daily cron for analytics refresh