Foreman automates tedious bulk work like triaging thousands of reviews, tickets or records. You describe the job in plain English. A light LLM (DeepSeek V4 Flash) plans it once into a pipeline of tools: filter, classify, group, dedupe, score, rank, calculate. Jev, a new fast decision model from TypeSafe, makes the per-item decisions in parallel and returns a calibrated confidence with each one. Any decision below the confidence bar you set gets one second opinion from the LLM. Math, counting and dates are always done in code. Every decision is logged with who made it and how sure it was, so you can click any result and see why.
Real run: 10,000 public app reviews became a ranked, deduplicated bug backlog for about $1.20, with 76% of decisions made by Jev. 1,000 reviews take about 1.5–2 minutes and cost about 10 cents.
Built with Python (async engine, FastAPI, SSE), Jev via Vercel AI Gateway, DeepSeek V4 Flash on Fireworks, and a React + TypeScript frontend.
Run it: see the README's "Run it" section in the repo (setup, keys, scripts/demo.sh, and a replay mode that needs no keys).
Run it locally
- git clone https://github.com/CloudCuckooLander722/EvoHacks-Submission-2026-Foreman && cd EvoHacks-Submission-2026-Foreman
- python3.12 -m venv .venv && source .venv/bin/activate && pip install -e ".[dev]"
- cp .env.example .env and add AI_GATEWAY_API_KEY (Vercel) and FIREWORKS_API_KEY
- scripts/demo.sh, then open http://localhost:5173 and try data/hero/demo_300.csv No keys? Press "Replay a recorded run" on the page to see a real recorded run.
Built With
- deepseek
- fastapi
- jev
- python
- react
- typesafe
- typescript
- vercel-ai-gateway
- vite
