human-AI system design

BOX BOX

Racing is becoming a conversation with AI. BOX BOX makes sure your engineer is always one step ahead.

An AI-powered race engineer for home sim racing. BOX BOX reads live telemetry, identifies critical moments, and delivers real-time voice strategies like a professional pit engineer—helping drivers react faster, understand their performance, and make smarter decisions on track.

See the Prototype Read the Research

Preview

Click to enter prototype

01 INSPIRATION

Home racing simulator problem

Rear view of a driver in a home racing simulator, holding the wheel while an F1 cockpit is on the TV

Over-the-shoulder view of a home sim-racing rig facing a first-person racing game

Side view of a driver in a bucket seat with wheel and pedals, watching an F1 game

Driver in a home racing cockpit facing a large screen showing a race track

Formula One (F1) strategy is the calculated plan a team uses to maximize their race pace and track position. In professional racing, communication between the driver and the race engineer is not merely a matter of reporting data. This aspect is NOT reflected in the at-home simulator.

F1 strategy engineers at AlphaTauri race operations and a Mercedes pit wall, monitoring telemetry and track maps through headsets

F1 strategy engineers analyze data, develop race strategies, and optimize decisions regarding tires, pit stops, and car performance; providing drivers with recommendations for a winning strategy is a judgment call that combines intellectual ability and experience.

02 RESEARCH

03 PROBLEM

04 WIN STRATEGY RESEARCH

Module	Key input	Judgement logic
Fuel	fuel_kg, fuel_mix, lap-history fuel deltas, remaining_laps, SC/VSC, MFD remaining laps	Blend observed + prior burn rate → finish margin after reserve → split save / neutral / push at 0.5 / 1.5 laps; no call if data missing or confidence is low
Tire	compound, age_laps, wear%, four-corner temps, track_temp	dt = lambda(alphaa+betaa^2) + (1-lambda)kappa*w + thermal; life = min(wear, age, performance budgets); overheat bands from T_hot / T_crit
Pace	lap_history plus fuel / tire / weather / damage / ERS derived fields	Drop pit/SC/yellow/traffic/out-laps → fuel-correct base pace → add tire/fuel/thermal/weather/damage/ERS → regress trend and project H laps
ERS	battery_percent, harvest/deploy J, deploy_mode, gap_ahead/behind, SC/VSC	SoC vs attack/defend floors + blend of observed net SoC change vs mode prior → permission flags + recommended mode
Weather	rain_percent, weather_code, air/track temp, remaining_laps	Rain intensity → wetness class → grip surface → drying/wetting rate → ETA until grip crosses g_inter / g_wet
Opponent	gap, last_lap, both cars’ tire derived fields, pit_stops, in_pit	Continuous score from gap + pace + tires + catch window + undercut → HIGH/MEDIUM/LOW → DEFEND/COVER/UNDERCUT/HOLD

Single Module Logic

I asked AI to organize the calculation parameters and strategies of each module into Python files.

Python _check function validating threat level, pace delta, undercut risk, and other strategy parameters

Python module files generated for each strategy model: __init__, __main__, model, params, and scenarios

scenarios.py tests the opponent strategy model. It loads YAML race cases, builds player and rival states, then scores threat. Checks compare threat level, pace, attack chance, catch window, undercut risk, and confidence with expected values, then print PASS or FAIL.

Overview Module Logic

advice - Complete analysis, 4 modules:
  1. Overall status (Pos / Gap / Tyre / Age / Wear / Trend table)
  2. Lap time comparison (Lap Time / Speed Comparison table)
  3. Tyre degradation forecast (Wear% / Cliff / Remaining laps / Rate)
  4. Tactical advice (L{n} | Command | Rationale)

voice - One-sentence tactical call (<20 characters, player car only)
  Example: "BOX IN 2 LAPS, front tyre 78%"