dmj.one

Work / Enterprise & Security / VSBS

No. 28 · In progress · Mobility · Research-grade

A booking flow with nobody on the line.

Booking your car for service is still a 1990s phone call. VSBS is the opposite. A LangGraph supervisor with a verifier on every tool call, a deterministic safety engine, a wellbeing score with ten peer-reviewed sub-scores, and a payment state machine that runs the same in sim and in production. Zero human operators in the loop. Currently in active development · CARLA-driven driving simulation work is blocked on a dedicated GPU host (10GB+ VRAM); local laptop hardware cannot carry the simulation load. If your organisation is willing to sponsor a GPU host for the simulation suite, or would prefer to engage me to integrate VSBS on your fleet platform, I would be glad to discuss the arrangement that suits you.

In active development Apache 2.0 DPDP 2025 native EU AI Act FRIA ML-KEM-768 hybrid WCAG 2.2 AAA
Sponsor a GPU host · become an infrastructure partner Engage me on your fleet platform See the loop
WIPIn active development
10+ GBGPU VRAM needed for CARLA
10Verified tools in registry
6LLM providers, one switch
0API keys for sim mode

Act I · The Problem

A 1990s call centre, in 2026.

Vehicle service is still synchronous, opaque, and built around the dealer's diary. Three problems compound, and customers absorb every one of them.

30%

Slots are lost

Dealers lose nearly a third of bookable slots to no-shows, mis-scheduling, and missed callbacks. Margin evaporates.

90 min

Customer time spent

The owner repeats the symptom three times, on hold, on email, in person. The mechanic still gets the wrong context.

0

Safety triage upfront

"Grinding when I brake" should never wait three days for an appointment. Today, it does. Nobody is checking.

Act II · The Promise

One conversation. Booked.

Type a symptom. The agent does the rest. Every step is auditable, deterministic where it must be, and shows its working.

  1. 01

    Owner speaks. VIN, OTP, symptom.

    The wizard captures intent in plain language. GET /v1/vin/:vin hits the real NHTSA vPIC database, free, no key.

  2. 02

    Safety engine fires first.

    A deterministic red-flag gate runs before the LLM. "Grinding when braking" returns severity HIGH and a "do not drive" message before any model speaks.

  3. 03

    LangGraph supervisor takes over.

    Each tool call gets verified by a cheap second model. Hallucinated tool calls fail the verifier and are dropped.

  4. 04

    Wellbeing scored. Slot proposed.

    A composite score with ten peer-reviewed sub-scores ranks dispatch options by stress, trust, and transparency, not just distance.

  5. 05

    CommandGrant minted.

    If the vehicle is autonomous, a signed capability token authorises the journey. Post-quantum hybrid envelope: ML-KEM-768 + X25519. Time-bounded. Revocable.

  6. 06

    Razorpay state machine settles.

    order → intent → authorise → capture → settle. Identical wire flow in sim and in production. Idempotent. Webhook-driven.

Act III · The Composition

Six things nobody ships together.

Any one of these is a research project. VSBS composes all six in a repo you can clone, build, and read end to end on a Sunday afternoon.

01 · Supervisor

LangGraph with a verifier chain

Every tool call is checked by a cheap second model for grounding before the result is trusted. Hallucinated calls are dropped, not executed.

02 · Provider-agnostic

One env var, six LLMs

Run on Gemini, Vertex Claude, OpenAI, Anthropic direct, or a scripted sim driver. LLM_PROFILE=sim needs zero keys.

03 · Sensors

Kalman fusion + SOTIF arbitration

Deterministic simulator. Distinguishes sensor faults from actual events. Aligned with ISO 21448.

04 · Prognostics

ISO 13374 health state machine

Refuses autonomous operation when a tier-1 safety-critical sensor is dead. Fails closed, by design.

05 · Privacy

DPDP 2025 native

Append-only consent log with evidence hashes. 72-hour breach runbook checked into the repo. FRIA documented.

06 · Crypto

Post-quantum hybrid envelope

Cloud KMS ML-KEM-768 + X25519 for long-lived secrets. Today's threats and tomorrow's quantum, both covered.

Act IV · The Stack

Production-shape, in one repo.

Every line is in the open. pnpm install, pnpm -r typecheck, six packages clean, full repo build, agent loop running in under a minute. Driving-simulation suite (CARLA) is paused pending dedicated GPU infrastructure.

  • TypeScript 5.7
  • Next.js 16
  • React 19 + Compiler
  • Hono 4
  • Bun 1.3
  • LangGraph 0.2
  • Tailwind 4 OKLCH
  • Zod schemas
  • Mem0 memory
  • Razorpay state machine
  • NHTSA vPIC
  • CARLA bridge
  • Cloud KMS
  • Trivy SBOM CI

Act V · Proof

It already runs.

SSE concierge endpoint

POST /v1/concierge/turn streams typed events: tool-call → verifier → tool-result → delta → final. Every arrow is real code you can read today.

CARLA bridge, optional

Replay mode runs in CI without a CARLA install. Live mode drives a Tesla Model 3 simulation at 10 Hz: 8 cameras, 4D radar, LiDAR, thermal, IMU, V2X, 96-cell HV pack.

L5 sensor dashboard

BEV view, perception event log, OOD score, UNECE R157 takeover ladder, MRM, capability budget, DMS gaze. Every channel a real production AV stack publishes off-vehicle.

Compliance evidence in repo

DPIA, FRIA, gap-audit, defensive publication, simulation policy, safety notice. Every architectural decision traces to a paper or a published standard.

Apache 2.0 with safety notice

Honest about what it is: research-grade reference architecture. Not a certified safety-of-life system. The repo says so on line one.

Defensive publication filed

Published as prior art so nobody can patent the loop and lock OEMs out. docs/defensive-publication.md.

If a service-advisor job can be agentic, so can your back office.

I build agentic workflows that ground every tool call, fail closed on safety, cite their sources, and run on a slow phone. Every architectural decision has a paper or a standard behind it.

Hire me Back to work