PitGPT Analyzer

Telemetry stint intelligence · pos-synced traces

How PitGPT works

Every stint runs through the same deterministic pipeline before a single word is spoken:

1Validate. The JSON is parsed and checked field-by-field with Zod. Tyre temperatures arrive as fixed-point integers and are divided by 10 to real Celsius on ingestion.
2De-spike. The speed channel is cleaned of corrupt samples (see section 2) so nothing downstream trusts a bad reading.
3Segment laps. Laps are split where track position wraps (pos crosses 1.0 → 0.0 and the lap counter ticks).
4Reject junk. Partial first laps, in-laps, out-laps and pit laps (speed dropping to a crawl) are filtered out — only clean racing laps survive.
5Per-lap metrics. Lap time, top / average speed and throttle & brake duty cycles are computed, all time-weighted.
6Map the track. Turns are found from steering geometry on the best lap, then projected onto every lap by position (see the Track Map panel).
7Localise the loss. The slowest lap is compared to the best to find exactly where time leaks away (section 3).

Recorder data is noisy and occasionally corrupt. Three defences keep the analysis honest:

De-spike filter (median-of-3)
A sample is flagged as a spike only if it is both a ratio outlier (more than 1.5× the larger neighbour) and large in absolute terms (≥ 20 km/h over its neighbours). It is then replaced by a linear interpolation of its neighbours. The dual test means a real glitch like 4021 km/h is caught while a slow pit crawl (0.1 km/h) is left alone. Every fix is logged to the console.
Fixed-point correction
Tyre temps are wire-encoded as integers scaled ×10, so they are divided by 10 immediately on ingestion to real Celsius.
Unstable frame rate
The stream rate drifts, so time always comes from real timestamps (dt = ts[i] − ts[i−1]), averages are time-weighted, and cross-lap comparisons use track position (pos) — never time — as the shared coordinate.

To keep the simulation realistic, the call is anchored to a specific instant in the race — the engineer only knows what happened up to that point.

1Pick the trigger lap. The slowest clean racing lap is chosen — that is where the biggest time loss lives.
2Find the problem turn. A position-based delta curve (worst lap minus best lap) is built; the turn with the largest jump in delta from entry to exit is the corner that cost the most time.
3Time the transmission. A real engineer speaks on the following straight, so the call fires 2.5 s after the driver crosses the line to finish that lap: trigger = last frame ts + 2500 ms. That is the [Lap X — radio call 2.5s after lap finish] plate.
4Ground the words. The measured loss report — with pre-computed directions (earlier/later, slower, hotter) — is handed to Claude, which turns numbers into a terse, human radio call. If no API key is present, a deterministic fallback call is generated from the same numbers.

There is no GPS or heading channel in the stream — only speed and steering. The schematic outline is dead-reckoned from the best lap:

1Distance. Each frame's travel is ds = v · dt — speed times the real time gap between samples.
2Heading. A kinematic bicycle model says heading turns in proportion to steering and distance: dθ = k · str · ds. Integrating gives the car's direction; integrating direction over distance gives its x, y path.
3Self-calibration. The unknown gain k (steer ratio × wheelbase) is solved, not guessed: a lap is a closed loop, so k is set to make the total heading come out to exactly one full turn (±360°).
4Anchoring the turns.Each detected turn's apex is dropped onto the nearest point of the outline, so the T1/T2/T3 list gets a spatial home. The problem turn is marked in red; clicking any dot zooms the traces to it.

It's an approximation — drift means the loop may not close perfectly, and near-symmetric stints degrade gracefully to a rougher shape. It's a schematic to orient by, not a survey map.

A low-confidence hint beats a confident hallucination, so the localization ships with a score computed from the data — never from the model. Three signals combine:

Concentration
What share of the whole-lap gap is explained by this one turn (time_lost ÷ total_gap). One corner owning most of the loss reads as a clear story.
Dominance
How far the top turn's loss stands above the next-worst turn. A runaway leader is trustworthy; a photo-finish between several corners is not.
Magnitude guard
Whole-lap gaps under ~150 ms sit inside stream noise, so they cap the score low no matter how the loss is distributed.

The three fold into a 0–100 score labelled High, Medium or Low (the badge on the radio panel). The label is also handed to Claude: on Lowit is told to say so plainly ("low confidence on T[X]") rather than assert a firm cause.

Drop stint .json

or click to browse

PitGPT · Engineer Radio

>>Awaiting a stint with at least two comparable racing laps…

Load a telemetry stint

Drop a recorder .json into the upload zone above to analyze laps, map turns and get a PitGPT engineer radio call.