Applied AI Engineer · Agentic Systems · Frontier LLM Work
A loop is an agentic system that runs itself. I build the harness around frontier models — research, plan, build, evaluate, self-heal — with gates that decide what is allowed to ship. The harness is the moat, not the model.
The loop, drawn
One continuous cycle. Nothing lands until EVAL passes — on a pass it promotes and the loop comes round again; on a failure SELF-HEAL repairs and re-enters at BUILD. No human in the paper-cut path.
Falsifiable proof
Six things built and running. Every claim carries a figure you could check — patterns, never product names.
A production Discord AI agent for a paying client: 80,000+ community messages mined into automated support & FAQ answering in the client's own voice, plus scheduled newsletters, live market data, and custom TTS voice cloning.
An enterprise semantic knowledge base: 102,000 vectors over 13,000+ documents drawn from 8 distinct sources, deduplicated and chunk-scored so answers cite the right passage, not the nearest-looking one.
A multi-agent orchestration kernel that replaced 200+ scattered cron jobs: 12+ concurrent agent identities, capability-gated approvals, and full audit logging on every effect.
Eval-gated promotion: a build is staged, scored against a rubric, and only promoted on a passing gate — the same discipline that keeps my own systems from shipping regressions.
vocax.ai — a multi-tenant assessment platform, live in beta and taking payments: occupational assessment scored against a weighted rubric, O*NET SOC crosswalk, org-tiered billing — built on the same gated, audited harness as the client work.
The reusable pieces, in the open:
Built, not published. Capability-gated approvals, eval-gated promotion and the orchestration kernel run in production and stay private. Publishing the harness would hand over the thing that makes the work compound.
I co-built and shipped Peak Finance, a seigniorage protocol on Metis Andromeda — live on-chain, real users, real money, and a team doxxed to the chain rather than anonymous. We wound it down when we couldn't finance it through the bear market, along with most of that cohort. Shipping systems where a bug costs someone money is not new ground.
The path here
Tristan Hazelwood
Coffs Harbour, NSW · working remote
I spent five years doing one-on-one, high-ticket client work at a crypto-education company — the last of them coaching people, hands-on, into actually adopting AI in their day-to-day.
That work taught me the thing most engineers learn too late: a model is only worth the system wrapped around it. The client who succeeds isn't the one with the smartest chatbot — it's the one whose workflow keeps running when the model has a bad day, when a source goes dark, when the input is malformed. The moat was never the model. It was the harness.
So I went independent and now build autonomous systems full-time. I design the loop — the research-plan-build-eval-self-heal cycle — and the gates that keep it honest: capability-scoped approvals, audit logs, eval-gated promotion. You tell me the outcome. I build the thing that builds it, and then hand you something that keeps working after I've gone.
How I work
The harness is the moat, not the model.
Loops and gates are the durable asset. Models get swapped underneath a stable harness — the system doesn't flinch when a better one ships next quarter.
Every effect is capability-scoped, approval-gated, and audited. Autonomy without gates is just a faster way to be wrong at scale.
Claims ship with numbers you can check. An eval you can't fail isn't a gate — it's decoration. I'd rather report a real blocker than a clean-looking lie.
See it run
One showreel and three teaching sessions from the course I run on AI engineering. Each one is the same loop in different clothes.
The hire page has the record behind these sessions and how to get in touch.
Notes from the loop
Published autonomously by the same pipeline that taste-gates this page. Freshest first.
Work with me
That list is where I've shipped, not where the edge is. I've built with an AI-native stack for two years, so the practical constraint is usually whether an outcome can be specified clearly — not whether a particular tool already appears on a CV.
Most questions about scope, stack and what I've actually shipped are answered faster by the agent in the corner — it reads from the same record I'd quote from, and it will tell you when something is outside what I've done.
When you already know what you want built, this lands directly with me. No pricing tables, no gig-platform links. I read every message myself and reply personally.
Or plainly: [email protected]