Problem Statement 1 · The Silent Co-Driver
ApexSignal transcribes team radio, scores acoustic tone, and checks every driver complaint against the driver’s own telemetry baseline — turning a subjective “rear’s moving” into measured evidence, connected lap-by-lap to what the car actually did.
Session
—
Driver
—
Incidents
—
Mode
connecting
Architecture
Three stages, each independently testable, connected only by the frozen JSON contracts in contracts/.
01
Team radio is transcribed with Whisper ASR and scored for acoustic tone/arousal independently of the words used. Every call becomes a structured record, not a raw clip.
POST /v1/radio/analyze
02
The reported complaint is checked against the driver's own recent laps at that segment, and against prior incidents in memory — never against a population average or a fabricated risk score.
POST /v1/incidents/evaluate
03
One screen: transcript, tone, complaint category, baseline evidence, historical match, and lead time — worded in interpretation-safe language, with an honest null state whenever the evidence isn't there.
GET /v1/incidents/{id}
Fixed complaint taxonomy · five categories, frozen
Tone model · real broadcast audio, real output, not a fixture
Real VoiceCLAP output
Most confidently CALM in the corpus
2024 São Paulo Grand Prix — Lando Norris, real broadcast
“If the weather stays as it does, we think it will be these tyres to the end.”
Most confidently ELEVATED_AROUSAL in the corpus
2018 Abu Dhabi Grand Prix, real broadcast
“These tires are not doing well! Copy”
Two real drivers, two real broadcasts, six years apart — the model wasn’t tuned on either. Same threshold used in production (AROUSAL_ELEVATED_THRESHOLD = 2.565, calibrated on 20 real human-labeled F1 radio clips) decides both. Run it yourself: services/radio_ai/tone_test/run_live_pipeline_demo.py.
Live Pipeline Walkthrough
Five real broadcast clips, fetched live from the API — same data the Pit Wall Console runs. Nothing here is a mockup.
Loading from API…
Circuit context · source geometry
Twenty-five circuit centerlines sampled from an open motorsport geometry database—not invented SVG loops. Twenty-one maps also rotate through complete, source-backed Formula 1 race recaps; map-only venues remain in the atlas when comparable F1 lap-order data does not exist.
25
source centerlines
21 race replays · 1,284 laps
Selected map · AUS
Melbourne, Australia · 5.299 km centerline
Lakeside semi-street loop with fast direction changes.
Geometry derived from TUMFTM/racetrack-database (LGPL-3.0). Historical grid, classification and per-lap order are bundled from the Jolpica/Ergast archive. Between-lap map motion is interpolated for the recap and is not presented as GPS telemetry. Centerline lengths are computed from source coordinates and may differ slightly from official homologated lap distances.
Validated Reference Replay
Loading replay data…
Interactive Replay
The validated replay is connected to the public ApexSignal API — click a pin, read the evidence, toggle Pit Wall View to see what changes.
Incident Inspector
Evidence
Loading evidence…
Human-in-the-Loop
Play a real clip. See the AI’s real call. Confirm it or correct it — the same workflow that grows the classifier’s training data internally.
“A lot of understeer creeping in. All speeds but also traction getting very poor.”
Your call
“Max, so specifically the issue is breaking turn 10 whilst on the curb, just for information.”
Your call
“Okay, copy, just keep the information coming.”
Your call
“Oscar confirming that tyre degradation is all thermal. Not worried about wear, not worried about graining on this tyre.”
Your call
Scope, Stated Honestly
A smaller, honest system beats a larger one that overclaims. Both lists below are enforced in code, not just in copy —test_contract_conformance.py checks the risk-score rule directly.
No lie detection
Tone/arousal is an acoustic model score, labeled as such. Never framed as detecting deception.
No diagnosis
"Reported phenomenon," never "diagnosed fault." ApexSignal never claims a confirmed mechanical cause.
No composite risk score
Every component — semantic similarity, telemetry similarity, tone confidence — stays visible and separate. Nothing gets collapsed into one number.
No recurrence prediction
ApexSignal flags a recurrence only after the driver reports it again by radio. It does not proactively watch telemetry in the background for a repeat before that happens.
Deferred to the roadmap, not this build
Recurrence monitor
A standing background process that watches telemetry independent of radio events was in the original concept. Cut for the MVP — recurrence is only assessed reactively, when a new radio report arrives.
ECHO LAP at scale
The presentation replay uses three contract-validated reference incidents. The same interface (semantic + telemetry similarity, category gate) can scale to a FAISS index over a full-season corpus.
The Mask
text_tone_disagreement (text vs. acoustic tone mismatch) is schema-supported but disabled by default pending real-clip validation gates — see services/radio_ai/VALIDATION_GATES.md.
Field Context
Cross-driver / cross-team context correlation is a roadmap idea from the project charter, intentionally out of scope for a one-driver, one-session MVP.
Materials
Team
AI Race Month · GrandPrix Hackathon @ Paytm · Problem Statement 1: The Silent Co-Driver
Workstream A
Jagrav
Data, telemetry, deterministic replay
Workstream B
Shashwat
Radio & language intelligence — ASR, tone, complaint classification
Workstream C
Tanish
Incident memory, evidence engine, core API
Workstream D
Mohit
Product UI, visualization, deployment
▌ApexSignal by Podium Finish · MIT License · GitHub