KaizenLap

The AI coach that turns telemetry into trophies

Comment

Inspiration

I love racing. I love Toyota.
The second I saw HackTheTrack, one question burned in my head: “What can I build that actually hands Toyota GR Cup drivers and engineers a real edge?”

Not another generic dashboard. I wanted an AI race engineer in the passenger seat that stays calm, looks at every lap, and immediately tells the driver the single biggest place to find time. That’s how KaizenLap was born.

What it does

KaizenLap is an AI driver coach for the Toyota GR Cup.

It takes official TRD race telemetry (7 tracks, 14 races for this deployment) and, for every driver and lap section, does three things:

  • Exact time lost vs. the theoretical best composite lap (fastest sectors in the field)
  • Why it’s happening — technique, track evolution, tires, or possible setup limits
  • Turns that into concise, concrete recommendation for what to try and improve in specific sections next session.

All of this is grounded in section times, consistency, and weather data, then translated by Gemma 3 into driver-ready coaching notes.

The result is a library of ~21,000 section-level recommendations that are specific to a driver, a section, and a set of conditions — not generic “you’re slow here” comments.

How I built it

The project has three main pieces: a data pipeline, an AI coaching layer, and the app surface.

Data & analytics

  • Load 144 telemetry CSVs (laps, sections, weather, positions).
  • Build a theoretical “perfect lap” per race from the fastest section times.
  • Compute per-driver deltas and consistency metrics, and correlate them with weather.

AI coaching (Gemma 3)

  • Feed Gemma 3 structured stats for one driver + section + context.
  • Use a strict template: facts → hypotheses → recommendations.
  • Pre-generate ~21k section-level coaching notes and store them as Firestore documents.

App & infra

  • Backend: FastAPI API that exposes track → race → driver → lap and reads from Firestore.
  • Frontend: React + Material UI with track maps, section overlays, comparison views, and a coaching panel.
  • Cloud: Containerised and deployed on Google Cloud Run with Cloud Storage for raw data.

Challenges I ran into

  • Turning cold numbers into warm, trustworthy coaching that actually sounds like a pro race engineer
  • Dealing with imperfect files and focusing on the cleanest race data for robust analysis.
  • Keeping Gemma 3 tied tightly to the telemetry so it does not guess beyond the numbers. (100% honest - zero hallucinated advice )

Accomplishments that I'm proud of

  • End-to-end pipeline: raw TRD CSVs → composite laps → 21,000 real coaching notes in <72 hours
  • A UI so intuitive that judges can pick any driver and understand exactly where time is hiding in under 60 seconds
  • Coaching notes that read like they came from a GR Cup series engineer, not an AI

What I learned

  • Consistency is often more decisive than a single peak lap.
  • LLMs can act like disciplined race engineers when they are tightly scaffolded around real telemetry.
  • Sector-focused views plus short notes are far easier to use than raw channels and huge tables.

What's next for KaizenLap

  • Add streaming feed and apply the same logic to live coaching and strategy.
  • Build long-term driver profiles across tracks, conditions, and stints.

Built With

Share this project:

Updates