Racing Telemetry Visualiser
A live AI pit crew for iRacing. RTV reads a car's telemetry 60 times a second, works out
the race state with plain code, and wakes five LLM agents (strategist, vehicle engineer,
spotter, driving coach and race control) only when something happens. They talk to the driver on one radio
channel, and every number they say is checked against the data before it is spoken. After
the race, a review replays everything the agents saw and said.
Open the live demo: the post-race
review of a real 42-minute GT3 race at Spa, replayed with the AI agents on. It runs in the
browser from static files, so there is nothing to install.
Why I built it
Sim racing tools tend to stop at one of two things: a database of laps and charts, or a
chatbot you can ask about them. A real race engineer does something harder. They watch the
race, keep quiet most of the time, and when they do speak the numbers are right.
That makes this a good test bed for a question that matters well beyond racing: how do you
put LLM agents in a real-time loop and still trust what they say? RTV's answer is to give
all the maths to deterministic code, let the models only explain and decide, and check
every figure they produce against the facts they were shown.
Highlights
- Real-time pipeline. Live shared-memory capture at 60 Hz or offline
.ibt import, stored
as DuckDB metadata plus per-lap Parquet, and served over REST and WebSockets. The variable
catalog is read from the telemetry header at run time, so it works for any car with no
per-car mapping.
- A race-state engine with no LLM in it. Fuel, tyres, pit window, gaps and around 30 kinds of
race event are computed by plain code. On the real Spa replay it averages about 0.1-0.2 ms
per frame, against a 16.7 ms budget at 60 Hz.
- Event-driven agents, not a chat loop. Agents wake only when an event matches one of
their triggers. In the recorded Spa replays a full 42-minute race used 25 to 44 model
calls, under a hard cap of 400.
- Every spoken number is fact-checked. A deterministic validator reads each figure and
its unit ("6 laps", "2.4 s", "P4") and matches it against the facts the agent was given.
A figure it can't trace is sent back for one repair, and if it still fails the agent
gives a grounded refusal instead of a guess.
- One radio channel with rules. Critical calls pre-empt advisories, a newer call on the
same decision replaces a queued one, and a call that is out of date by the time it would
air is tagged late or dropped.
- Measured, not just demoed. Every run is written to a race journal. A run-comparison
script, replayable eval cases cut from a real race, an LLM judge on a second model family
(Gemini) and a weekly scheduled eval track whether a change made the agents better or
worse.
- Native desktop app. RTV Desktop (PySide6/Qt) embeds the server and is the main UI, with
local push-to-talk questions (faster-whisper, no audio leaves the PC).
- Tested offline. Over 1,400 tests run with no game, no network and no API key, on
Python 3.11-3.13 in CI, plus a job that runs the real 162 MB race file end to end.
What it looks like
| Live pitwall | Lap analysis |
|---|
| |
Strategy, fuel margin, pit window, tyres and the radio during a replay of the Spa .ibt. This capture has the agents off, so it shows only what the deterministic engine produces. | Two laps compared on lap distance, with a time-delta curve, the track map and the corner-by-corner findings that every AI layer is grounded on. |
How it works
flowchart LR
A[iRacing shared memory<br/>60 Hz] --> I[Ingest]
B[.ibt recording] --> I
I --> S[(DuckDB metadata<br/>+ per-lap Parquet)]
I --> E[Race-state engine<br/>detectors, no LLM]
S --> F[Lap findings<br/>deterministic coaching]
E -- race events --> O[Orchestrator<br/>triggers, cooldowns, budget]
O -- template calls --> R
O --> G[Agents on Kimi<br/>strategist, engineer,<br/>spotter, coach, race control]
F --> G
G --> V[Fact check<br/>every figure vs the data]
V --> R[Radio feed<br/>priority, supersede, late/expired]
R --> U[Desktop app, web views,<br/>voice]
E --> J[(Race journal)]
G --> J
R --> J
J --> P[Post-race review<br/>and evals]
This is the beginning of the README. Read the rest at the original. Read the full README ↗
