Oracle × Scuderia Tor Vergata hackathon · 2026
AURIGA
An AI race engineer: it reads the telemetry, cites the rulebook and simulates the race before recommending a pit stop.
live demo Simulated console + original Strategy Lab
The race is the simulated one provided by the hackathon. The demo exposes the Strategy Lab with the original engine on the server; agents, prompts and Oracle services remain private.
Session live · lap 23 of 50
Race engineer
The model never does maths in its head: numbers come from typed tools, and every answer cites its source.
- Lap
- 23/50
- Front-left wear
- —
- Pit window
- —
- Cost of the answer
- —
- Pick a question to see which specialists are called and what they return.
- FL · front left
- FR
- RL
- RR
| Measure | Dossier | Limit | Result |
|---|---|---|---|
| Chord | 308 mm | ≤ 300 mm · §3.2 | needs revision |
| Width | 940 mm | ≤ 1000 mm | compliant |
| Height | 1150 mm | ≤ 1200 mm | compliant |
“Box, box on lap 25. Front-left is fading: go easy on entry to the Tornantino. Everything else is fine.”
Abnormal FL degradation from lap 17 (+18 pp at lap 23). Stop planned for lap 25, estimated loss 36 s.
Lap 23 at Vallelunga: strategy updated, pit stop coming up. 🏁
Original interface · original engine on the server
Strategy Lab
Change fuel, consumption per lap and the two drivers’ pace: the project’s Monte Carlo simulator runs thousands of races and returns the probability of finishing, expected score and the best driver order.
Up to 5000 simulations per request. No AI agents involved: this is the pure calculation module. Interface in Italian.A — The problem
In an endurance race, decisions happen in seconds: pit now or in three laps? Is that part legal? What do I tell the driver? The data exists — telemetry, rulebook, technical dossier — but it is scattered, and nobody can read it all while the car is on track.
B — What I built
For Oracle’s AI Race Engineer Challenge with Scuderia Tor Vergata I built AURIGA: a supervisor that routes every question to five specialists — telemetry, rules, strategy, communications and team memory. Each works with its own tools, and the interface shows live who is doing what. What comes back is a single recommendation with its sources: the laps analysed, the rule article, the simulation that was run.
C — How it works
- 01
The question
The pit wall asks in plain language: “Should we box?”
- 02
Routing
The supervisor picks the specialists it needs and runs them together.
- 03
Tools, not hunches
Wear per lap, pit windows, component checks: numbers come from typed tools, never from the model’s imagination.
- 04
Simulation
A Monte Carlo engine runs thousands of races with the competition’s scoring rules and compares strategies.
- 05
The call
One answer, reasoned and cited: “box between lap 24 and 26”.
D — The system
A simplified map of the components: not the technical documentation, but it shows how the parts talk to each other.
E — Why it works
The model never does maths in its head
Strategy and fuel always go through a simulator using the real scoring formulas.
Every verdict cites the rule
A component check returns the measurement, the limit, the rule section and the outcome.
Everything visible live
Agent constellation, tool calls and costs stream in real time.
Honest about scope
Built and tested on the hackathon’s simulated race, not on real track data.