James Westover

Where I've shipped

Safety-critical AI · ISO 9001 audit-ready

Cloud architect for the HAD trilateral — BMW · Daimler · Audi.

• Architected the cloud infra for HERE's Highly Automated Driving program — three OEMs, simultaneous sensor pipelines, autonomous-vehicle map ingest at scale.
• Owned ISO 9001 documentation, workflow diagrams, and a week-long external audit. Most AI engineers have never been through QMS audit — for healthcare ISO 13485 the discipline transfers directly.
• First-principles bandwidth math: LTE uplink could not carry AV sensor data — shipped physical drives to collection centers. Inspired by Jim Gray: "never underestimate the bandwidth of a FedEx truck."

GDPR at billion-trip scale

Engineering Manager · Location Platform · Year 1 at Uber.

• Led GDPR compliance push across Uber's location data platform — right-to-be-forgotten, data deletion, audit trails for every trip, every rider, every driver.
• Cross-team coordination across Maps, Identity, Riders, Drivers, Legal — at a company that was simultaneously fighting regulators on multiple continents.
• One of many production initiatives that year — Shadow Maps GPS refinement and several others shipped in parallel, same team, same year.

$30M / year fraud — and the arrests that followed

Production ML versus organized adversaries.

• A few months back at Uber as Staff Eng, built fraud-detection systems that captured $30M in annual savings. ML + signals engineering across rider/driver/trip telemetry.
• The vivid part: photos of tables of phones in Asia — phone-farm operations exposed by the model. Real-world consequences: arrests in Miami.
• Hard numbers, hard adversaries — fraud detection at marketplace scale is the same shape as clinical-accuracy detection: false negatives are expensive, false positives erode trust.

From 45 minutes to 12 — 150 devs unblocked

CI infrastructure on Kubernetes pod runners.

• 1M+ LOC monorepo · 17K+ tests · ~150 engineers. CI builds were 40-50 minutes — every PR burned a context switch, every developer felt it.
• Rebuilt the test infrastructure on Kubernetes pod runners with sharded distributed execution. Result: 45min → 10-12min. Across the org, that compounds.
• Concurrent: led the Python 2 → Python 3 migration for the full Postmates monorepo. (Same pattern repeated at Uber: ~200 libraries across 4,000+ microservices.)
• Integrated Envoy for blue-green deploys and OpenTelemetry for observability.

A working agent platform — mine

4 daemons · multi-LLM fan-out · knowledge graph with continuous compaction.

• Hive · Bahnhof · Ausweisamt · Tunnelmeister — four-daemon architecture managing persona sessions, transport, identity, and tunnels. Long-lived agents that don't lose state.
• Forge multi-LLM fan-out: Claude / Gemini / local Ollama / Hush — same prompt to all, compare outputs, pick the right one for the job.
• Knowledge Graph on Postgres + pgvector with continuous compaction (PG triggers + Forge fan-out for dedup). Persona memory that scales.
• 2 iOS apps with paying subscribers in the App Store. The platform isn't a slideware.

This deck wrote itself.