About the Project — Race Oracle

Inspiration

Every race has a story hidden beneath the lap charts. Tiny steering twitches, sudden drops in speed, brake pressure spikes, and G-force shifts reveal the real narrative that fans and commentators usually never see.

We wanted to build something that uncovers that invisible layer — a system that can read the race like an engineer and explain it like a commentator. A tool that doesn’t just show what happened, but why it happened.

That idea grew into Race Oracle, an AI engine that reconstructs the entire race using telemetry, timing, weather, and sector data.

What We Learned

Working on Race Oracle taught us how much complexity hides inside motorsport data. We learned:

How raw telemetry signals reflect real driving behavior: steering variance, aggression, smoothness, consistency.
How to merge six different CSV formats into a unified structure.
How to design performance metrics such as throttle smoothness, brake spike detection, and lateral G efficiency.
How to identify key racing events from data instead of video.
How to prompt LLMs to produce reliable engineering-style explanations without hallucinations.

We also deepened our understanding of how time loss actually works in racing. This helped us quantify “mistakes” in a consistent, data-driven way.

How We Built It

Race Oracle is built on three major components:

1. Data Processing & Cleanup

We ingested and standardized:

Telemetry (speed, throttle, brake, G-forces, RPM, steering)
Timing & scoring (lap times, sectors, best laps)
Race control & metadata (positions, gaps, classes)
Weather & conditions

From this, we produced structured datasets:

telemetry_clean
lap_times
driver_metrics
event_logs
race_summary.json (the master file)

2. Analysis & Detection Engine

This system extracts:

Driver performance metrics
Pace evolution across laps
Steering smoothness & aggression profiles
Brake spike and lockup detection
Understeer / oversteer moments (via mismatch of steering angle & lateral G)
Race turning points and time-loss events

3. AI Storytelling & What-If Simulator

We built a multi-stage LLM pipeline that generates:

Full race summaries
Driver storylines
Race turning points
Performance debriefs
Setup recommendations
Alternate-universe “what if” simulations

Users can type questions like:

“What if Driver 26 didn’t slide wide on Lap 4?”

And the system describes how the race outcome would shift.

Challenges We Faced

Telemetry sampling rates were inconsistent and required interpolation.
Some laps contained impossible values and needed intelligent filtering.
Distinguishing real driver corrections from sensor noise required smoothing and modeling.
Getting the LLM to stay grounded in data instead of hallucinating was surprisingly hard.
Merging sector data, weather info, and race results into one timeline required custom logic.

Each challenge made the system more robust and the insights more accurate.