TRIDENT

TRIDENT is a deterministic motorsport intelligence core that ingests telemetry, models laps, detects anomalies, and builds evidence-driven driver insights, with AI only narrating results.

Comment

🏎️ TRIDENT - Toyota Race Intelligence & Driver Engineering Toolkit

🚀 Inspiration

Motorsport produces massive amounts of telemetry, but translating those signals into clear, engineering-grade driver insight usually requires hours of manual analysis.
I wanted to build something deterministic, transparent, and mathematically grounded an engine that extracts real evidence instead of relying on LLM speculation.

My goal was to create a backend intelligence system that:

  • Understands drivers through structured telemetry
  • Generates repeatable, evidence-backed insights
  • Scales instantly to any future Toyota dataset
  • Lays a foundation for real-time race engineering

The COTA dataset became the perfect proving ground.

🧠 What it does

TRIDENT is a driver intelligence engine that turns raw Toyota COTA telemetry into structured, evidence-supported coaching insight.

It delivers:

  • SIWTL Ideal Lap Modeling (realistic best-possible lap)
  • Predictive Lap-Time Modeling with XGBoost + feature attribution
  • DPTAD Anomaly Detection for mistakes, lockups, degradation, pacing issues
  • Driver Behavior Profiling (aggression, smoothness, hesitation)
  • Consistency, variance, and pace metrics
  • A unified evidence graph for each driver
  • AI-narrated coaching cards generated only from verified data

The frontend is only a viewer; the intelligence is fully backend-driven.

🛠️ How I built it

I engineered TRIDENT as a modular, deterministic pipeline where every insight is computed, not guessed.

1. Data Architecture

  • Normalized 100+ COTA files (laps, sectors, telemetry, weather)
  • Built a canonical Parquet store using DuckDB for analytical performance

2. Feature Engineering

I engineered 117 features, including:

  • brake_spike_count
  • throttle_variability
  • sector_delta_rate
  • degradation_slope
  • consistency_score
  • smoothness, aggression, hesitation indices

3. DPTAD — Dual-Path Time-Series Anomaly Detector

Custom-designed anomaly system:

  • Fast Path: instant mistakes (spikes, late braking, early lifts, slow exits)
  • Slow Path: degradation, brake fade, pacing decay
  • Every anomaly is clustered and labeled with metadata

4. SIWTL Ideal Lap Engine

  • Computes a realistic, not theoretical, ideal lap
  • Finds statistically achievable gains
  • Avoids unrealistic best-of-each-sector stitching

5. Predictive Lap-Time Model

  • XGBoost trained on 19k+ laps
  • Feature importance revealed peak brake pressure as top contributor
  • Integrated directly into TRIDENT Core

6. Evidence Graph Compiler

I combined:

  • ML predictions
  • anomalies
  • sector deltas
  • degradation patterns
  • behavior signatures
  • consistency metrics

All of this becomes a structured evidence pack per driver.

7. Lightweight AI Narrator

The AI layer only narrates results.
Intelligence = deterministic.
LLM = explanation only.
This ensures zero hallucination and engineering transparency.

⚠️ Challenges I ran into

  • Aligning lap, sector, and telemetry clocks into one timeline
  • Creating robust behavior metrics without subjective assumptions
  • Ensuring the system stays deterministic, not LLM-driven
  • Achieving dataset-agnostic design for future Toyota races
  • Managing irregular race laps and missing telemetry segments

🏁 Accomplishments I’m proud of

  • Building a fully deterministic motorsport intelligence core
  • Designing DPTAD specifically for racing telemetry
  • Engineering SIWTL as a realistic ideal-lap estimator
  • Producing insights grounded entirely in evidence
  • Achieving a track-agnostic backend architecture
  • Delivering professional-grade driver analysis automatically

📚 What I learned

  • Motorsport data rewards structure more than model complexity
  • AI must be constrained by physics and evidence
  • A small reasoning model with strong evidence beats a large model with none
  • Drivers don’t need vague coaching they need measurable, proven improvements
  • Robust pipelines make everything else easier

🔮 What’s next for TRIDENT

Even though this project is submitted under the Driver Insights category, the underlying system is built for more.

1. Real-Time Race Mode

  • Streaming telemetry
  • Instant anomaly detection
  • Live ideal-lap delta
  • Real-time coaching

2. Automatic Multi-Track Scalability

  • Adapt instantly to any new track’s structure
  • No manual reconfiguration

3. Corner-Level Micro Models

  • Per-turn behavior modeling
  • G-force analysis
  • Micro-anomaly detection

4. Race Engineer Copilot

Conversational analysis:

  • “Explain this anomaly”
  • “Compare Driver A and Driver B at Turn 11”
  • “Find the biggest time loss in the final stint”

5. Strategy & Simulation Layer

  • Under/overcut prediction
  • Tire degradation modeling
  • Optimal pit window forecasts

TRIDENT is built to evolve into a full-scale, mathematical backbone for Toyota’s next-generation driver development and race engineering ecosystem.

Built With

Share this project:

Updates