Everything here is built, measured, and public.

The nation's public AV crash dataset runs on self-reported automation labels, and nobody independently checks whether they're internally consistent.

Solution: a three-signal audit that never overclaims. Explainable heuristics, a LightGBM disagreement model at a deliberately chosen operating point, and a fully local LLM narrative read are fused into a short, human-reviewable queue, framed as review candidates, not verdicts, because the data can't support more.

• 4,589 cleaned records from NHTSA's Archive 2021–2025 snapshot (1,899 ADS + 2,690 ADAS)
• Operating point chosen for precision over recall: 0.912 / 0.378 on synthetic label flips at threshold 0.59
• Every stage's output is a tracked, machine-readable artifact under results/

Off-the-shelf detectors look fine on aggregate metrics while missing exactly the road users who matter most: small, distant, or poorly lit pedestrians and cyclists.

Solution: turn that failure mode into an auditable benchmark. Measure a COCO-pretrained YOLO baseline stratified by lighting, weather, and object size, not just mean AP, fine-tune on VRU-heavy data, then pick per-class thresholds with an explicit miss-cost argument. Ships a model card and an error report that says where it still fails.

• Fine-tuning lifted aggregate precision / recall from 0.396 / 0.248 to 0.685 / 0.335
• The honest headline: night-person and small-object recall remain the product-critical gaps, that finding is the point
• License-aware by design: no BDD100K pixels or restricted weights ever committed

Regulators give operators as little as 24 hours to report a qualifying safety event, and the deciding-what-matters step still happens in spreadsheets.

Solution: a config-driven agent that reads any tabular safety-event feed, classifies severity with deterministic rules plus a tool-calling LLM, and drafts the plain-language summary a reviewer actually needs. Fully local by constraint, not preference, safety data is sensitive, so everything runs on-machine through Ollama and no byte ever leaves.

• One YAML config per domain, identical code: proven on NHTSA AV crash reports and OSHA severe-injury reports
• Eval harness scored against 60 hand-reviewed ground-truth labels, the agent is measured, not vibes-checked
• Zero-dependency review workbench with rules-vs-LLM reasoning side by side

Cities wait four to six months and five figures for an intersection study that mostly assembles data they already have.

Solution: crash data in, draft engineering report out. Triage a corridor in under thirty seconds, open a full investigation, MUTCD signal warrants, volumes, crash clustering, and download a PE-reviewable draft PDF. Built as a real product, not a demo: marketing site, multi-city data lake with live FARS and California CCRS ingest, and a written go-to-market playbook.

• MUTCD Warrant 1 and Warrant 3 math plus all-way-stop screening, unit-tested against statute logic
• Liability handled like it matters: the DRAFT, pending PE review disclaimer is enforced on all six output surfaces, including the PDF watermark
• Scope and out-of-scope defined in writing before the first line of code