↗
From question to evidence
Choose an example above to reveal the real route, sources and response time.
ENGINEERING SHOWCASE · LOCAL ENVIRONMENT
See the intelligence.
Prove the operation.
Follow an answer, evaluate a release, and watch the gateway recover.
ConnectingSimulated vehicle · local control plane
Independent portfolio project
Independent portfolio project
REQUESTS IN MEMORY0Last 100 · cleared on restart
OBSERVED RESPONSE TIME—Latest pipeline time · excludes network & audio
ACTIVE CONFIGURATIONDefaultGateway priority routing
RELEASE ASSURANCEDemo onlyHuman and acoustic review separate
OBSERVE
Every answer has a path.
Run a question here or ask from the driver view. Only executed stages appear.
Questions and answer previews stay in this process’s memory. No audio is recorded. A trace shows provenance, not proof that every generated claim is correct.
EVALUATE → ACTIVATE → ROLLBACK
Earn the next release.
Create and evaluate both profiles to compare their frozen replay results side by side.
Default
Temperature 0.2 · concise output limit 110 tokens.
→
Candidate
Temperature 0.1 · concise output limit 96 tokens.
→
Local activation
Changes model requests in this assistant. Direct facts still use code.
Local release history
No releases yet.
Gate: all 21 frozen turns pass, the 3 model answers use the selected model without cache hits, and each model turn finishes within 35 seconds. Evidence is bound to source and expires for activation after one hour. This is an automated demo gate, not automotive safety approval. Saved LoRA artifacts are not deployed by this panel.
CONTROLLED FAULT INJECTION
Make failure visible.
Inject a timeout and watch the real routing algorithm fall back, skip the failed provider, and recover.
PRIMARY⇢GATEWAY⇢BACKUP
The demo creates isolated providers and circuit state.
The driver’s model connection and vehicle state keep running.
An injected 80 ms timeout trips a one-failure circuit threshold; after a 1-second cooldown the primary recovers. This exercises the production router with test providers. It does not measure real cloud outage recovery or model answer quality.