Gabriel Goodhart
AI Governance & Applied AI
- 38
- AI systems built
- 25
- live in production
- 2
- in the care sector
- Jul 2023
- building with AI since
Summary
I build AI systems and the governance that keeps them honest: multi-model routing, non-LLM scoring layers, dataset-backed evals that gate the release, human review checkpoints, and safety that fails closed. 38 systems since July 2023, 25 running in production, 2 of them in the care sector where PHI handling under HIPAA is a first-commit design constraint rather than a retrofit. I own the whole line — the routing, the eval harness, the review gate, and the thing that pages me when it breaks — which is why the controls I argue for are ones I have also had to live with. Every demo runs on sample data only; no client or personal data is exposed anywhere.
Selected Work
AILIVE
Shipped Groq-tutored exam-prep SaaS. Includes Stripe billing, SM-2 spaced-repetition scheduling, bilingual (EN/ES) content, and deterministic localization pipeline (~3,400 items, 5-stage QA gate, resumable, cost-capped).
Groq · Next.js · Supabase · TypeScript · Stripe · SM-2 algorithm
One blog topic fans out to Facebook caption, Instagram caption + hashtags, and email subject + body via a single Groq call. Nightly cron. Shipped as admin tool in production.
Groq · Next.js · social publish API · nightly cron
Multi-Model QA Cascade
AI
One item routed to Ollama + Groq + Codex in parallel; non-LLM scoring layer picks winner; human-in-the-loop review gate before SQL commit. $0 to run locally.
Ollama · Groq · OpenAI/Codex · deterministic scoring · human-in-loop · SQL
agent-gate — Safety Fence
AI
Deterministic guardrails for AI agents. 53 tests, fail-closed design: any uncaught edge halts the agent rather than letting it proceed. Built as a standalone reusable safety layer.
deterministic rules · fail-closed · 53-test suite
Vision Factory — Dual-Model Consensus
AI
GPT-4o generates assets, Gemini verifies. Fails closed if centroid drift exceeds 85px. Coordinates large consistent asset-generation runs via a spec-driven pipeline with a status board (todo → generating → verify → placed).
OpenAI GPT-4o · Gemini vision · consensus gate · asset pipeline
LIVE
Address → ranked candidate matches from 25+ sources. Weighted issue-scoring engine with auditable per-issue breakdown. Determinism-audited: same inputs always produce the same ranked output.
Next.js · TypeScript · determinism audit · 25+ public record sources
blood-suga — Vision Meal Analysis
AIDEMO
Live Groq text call per dish returns carb/calorie range; eval loop computes MAPE vs USDA-derived ground truth in real time. Production path uses llama-4-scout vision. Offline fallback labeled honestly.
Groq · llama-3.3-70b · dataset-backed eval · MAPE · offline fallback
Capabilities & Stack
Models orchestrated
Claude
OpenAI / Codex / GPT-4o
Groq
Gemini vision
Local Ollama
Stable Diffusion
Infrastructure
Next.js
TypeScript
Supabase Realtime
Playwright
Tailwind
Vercel
AI / Pipeline
Deterministic pipelines
Dataset-backed evals
Non-LLM scoring
Multi-model routing
Fail-closed safety
MAPE / eval loops
Systems shipped
SaaS (Stripe + SM-2)
Publish automation
Human-in-loop workbench
Agentic safety fence
Schema-driven intake
Unit-economics engine
How I Work
Deterministic first. LLM output is wrapped in rules, scoring, and gates so the system behaves predictably even when the model does not.
Verifiable by design. Evals run against real labeled data; MAPE and pass/fail rates are visible in the demo, not buried in a notebook.
Fail-closed safety. Any uncaught edge halts the pipeline. No silent fallback to "probably fine." agent-gate is a dedicated reusable layer for this.
Human-in-the-loop. Operator review gates are a first-class component, not an afterthought. audit-kit ships a typed JSON workbench the next stage reads back.
No client data in demos. Every demo tile on the portfolio runs on sample data only. Real products operate under separate, isolated data paths.