Safe LLM Driving Intent Platform
Design a safety-bounded service that proposes high-level driving intents without giving a probabilistic model control of vehicle actuation.
Problem statement
Design a platform that lets an autonomous-driving stack use an LLM for optional high-level maneuver proposals while remaining safe when the model hallucinates, times out, or is unavailable. For this exercise, assume each vehicle supplies versioned structured scene and route context rather than raw sensor streams.
The LLM returns a typed proposal with a maneuver, confidence, rationale, and references to the context it used.
A deterministic safety controller is the only authority that may approve and execute actuation; it must validate collision risk, legality, context freshness, and confidence, and it must continue with the verified conventional driving stack whenever a proposal is late, malformed, or rejected.
Include privacy-filtered auditability for incidents, but exclude model training, raw-sensor retention, and low-level actuator-controller design.
What to build
- Accept versioned structured scene and route context and return a typed high-level maneuver proposal with confidence, rationale, and traceable context references.
- Validate every proposal against deterministic collision, legality, freshness, and confidence rules before it can influence the driving plan.
- Continue driving through the verified conventional stack when a proposal is unavailable, late, invalid, or rejected, without waiting for the LLM.
- Record privacy-filtered proposal, validation, model-version, and fallback decisions for incident analysis.
At scale
- For this exercise, assume a fleet of 100,000 vehicles with 10,000 concurrently active vehicles.
- For this exercise, assume an optional LLM proposal has a hard 250 ms deadline; missing that deadline must not delay the conventional driving stack.
- For this exercise, assume the conventional safety stack remains locally available during regional model-service or network outages.
- For this exercise, retain privacy-filtered decision records for 30 days and do not retain raw sensor streams in this platform.