Jev Grand Prix
Local F1 sim where Jev picks racing line and pedals while code steers at 120 Hz; lap-by-lap race engineer learns corner limits between API calls.
Overview
Jev Grand Prix is an open F1 toy from enoyola. A Python server turns car telemetry into engineer prose, batches three System One questions (line Choice, pedal Choice, trouble Noul), and lets browser physics steer toward the chosen line at 120 Hz. Between laps, code compares sector notes and asks Jev once per corner whether to push, hold, or back off on a grip-percent ladder documented in the README.
Problem: Racing AI that steers the wheel directly from slow model replies weaves off track; F1 speeds cover twenty meters before an answer returns.
Built for: Developers learning the goal-versus-actuator split for real-time sims and hobbyists who want a local Jev race engineer.
First indexed on Jev Directory: 2026-09-21
Creator and team
- Name
- enoyola
- Handle
- @enoyola
How Jev is used
- Role in the product flow
- High-level racing line and pedal selection; separate lap-level corner pace planner
- Primitives
- ChoiceNoul
- State in
- Speed, track position, corner context, and natural-language sector notes generated from lap timing (see server.py and README tables).
- Decision out
- Target line across five lateral buckets, pedal mode among throttle/brake steps, and trouble probability; per-corner faster/same/slower plan between laps.
- Browser sends telemetry several times per second
- Server composes engineer text and batches three questions per request (~0.27 s cited)
- Client steers and applies brake-by-wire toward targets at 120 Hz
- At lap end, code summarizes sectors and asks Jev for per-corner pace adjustments
The README is explicit that Jev picks goals while code executes them, matching the Pokemon and Minecraft patterns TypeSafe cites. Noul powers the Trouble meter so the model can express leave-the-track risk without free text. The race-engineer ladder is a teaching tool: memory lives in deterministic lap analytics, while Jev only judges the next incremental pace change per corner. You can race Jev or drive yourself with arrow keys on localhost:8765 after uv run server.py.
Sourced performance claims
- Typical Jev answer latency about 0.27 s; direct wheel steering caused nine off-tracks in the first test lap.Source: github.com/enoyola/jev-grand-prix README
- Eight-lap test improved best lap from 59.2 s (standing start) to 50.9 s with real Jev calls.Source: github.com/enoyola/jev-grand-prix README lap table
Features and stack
Features
- Local browser client plus Python server (uv run server.py)
- Three batched questions per driving decision
- Optional human driver with camera toggle
- Race engineer panel with per-corner history
- Documented eight-lap improvement curve with real API calls
Stack
- Python
- uv
- Browser client
- TypeSafe System One API
Pricing: Requires your TypeSafe API key; README uses console.typesafe.ai early access.
Links
FAQ
- Why not steer the wheel directly from Jev?
- README reports weaving and nine off-tracks when Jev commanded the wheel; line plus pedal targets with fast local steering fixed clean laps.
- Which primitives are used?
- Choice for line and pedals, Noul for trouble probability, per README question table.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.