How does VoltKeep's AI analysis work?
A large language model analyzes your driving data (speed, acceleration, regen braking strength, ambient temperature, elevation change, etc.) and produces a combined score across driving efficiency, regen utilization, and speed patterns — together with concrete efficiency recommendations in natural language. The monthly AI report summarizes a full month of driving and recommends next-month improvements. The current backend is llama-3.3-70b-versatile hosted on Groq. (AI output is currently in Japanese; English output is on the roadmap alongside the English UI release.)
What inputs the AI sees
Analysis quality is bounded by input quality. VoltKeep aggregates the following signals from Tesla Fleet Telemetry and feeds them to the model:
- Speed profile — second-level vehicle speed, plus average / max / distribution
- Acceleration — strength of accel and decel, plus frequency of hard accel/braking
- Regen braking — kWh recaptured by regen, share of brake-pedal-assisted decel
- Power consumption — Wh/mi (or Wh/km) over time, broken down by drive mode
- Environment — ambient temperature, elevation change, HVAC use
- Charging patterns — start-of-charge SoC, charging speed, location classification
We don't feed the model anything Tesla doesn't provide (subjective driving-style notes, etc.). Outputs are constrained to what's defensible from the objective signals.
How the AI Driving Score is calculated
The AI Driving Score is a single 0–100 number combining three axes:
- Driving efficiency — how close your Wh/mi is to the baseline for the same conditions (temperature, elevation, vehicle)
- Regen utilization — when decelerating, how much you rely on regen vs. the brake pedal
- Speed optimality — how much highway time you spend in the most efficient 50–62 mph (80–100 km/h) band
Alongside the score, the AI proposes 1–3 concrete next steps in natural language (e.g., "Ease off regen on long downhills and supplement with the brake pedal — this can improve efficiency by 3–5%").
The monthly AI report
At the start of each month, the previous month's driving is rolled up into a single report:
- Total miles, average efficiency, score trend
- Repeated-route detection with hints for further optimization
- Efficiency by temperature band (e.g., how much winter hurts your range)
- Behavioral suggestions to try next month (preconditioning, cruise speed adjustments, etc.)
Backend details (for transparency)
The AI analysis currently runs on llama-3.3-70b-versatile hosted by Groq, accessed through an OpenAI-compatible API. We do not pass Tesla credentials or personally identifying information to the LLM — only the hashed VIN and aggregated data points.
The backend is implemented as an OpenAI-compatible API wrapper, so we can switch to other LLM providers (Gemini, OpenRouter, Cerebras, etc.) without code changes. This keeps failover straightforward when uptime matters.
Let an AI analyze how you drive your Tesla.
Connect for free, then unlock the AI Driving Score with Pro at $2.99/month.
Get it on the App Store