Agent harnesses
File-driven agents with roles, memory write-back and human-approved cleanup. Every model output is validated by code before it becomes a finding.
Auckland, New ZealandAI Engineer · Full-stackPortfolio, 2026
I build AI systems where code decides and the model drafts.
Most of my work sits at the seam between a language model and a business that cannot afford to be wrong — an earthworks contractor's invoices, a grower's harvest targets, a factory's downtime, an immigration adviser's case file. I design the harness around the model: the contracts it must satisfy, the checks that run before anything is shown, and the record of what was never read.
§ 01 — Craft
What I am good at
LLM application architecture
Role-based model routing, vendor fallback, rate limits, structured output and evidence-cited answers. Models evaluated per task, not by reputation.
Durable workflows
Long-running AI pipelines on Temporal: retries, timers and idempotent side effects, so a crashed render or a slow model never leaves half-written state.
Systems integration
Joining tools a business already runs — Xero, CostX, EROAD, packhouse exports, AXIS cameras — on a key that actually holds, with idempotent imports.
Full-stack product
Python & FastAPI, PostgreSQL with row-level multi-tenancy, Next.js & TypeScript. Interfaces that read as real operations software, not templates.
Shipping & operations
Cloudflare, Vercel, Fly.io, AWS and Oracle Cloud VMs; MQTT ingestion from the edge; CI with tests that verify something real.
“What was not read or judged is reported as a gap — never as clean.”
A rule from Cairn, and from most of what I build
§ 02 — Selected work
Seven projects
SiteLedger Field to invoice.
A job-management and billing platform for earthworks contractors that follows one story end to end: work happens on site, is captured on a phone, becomes priced invoice lines, and syncs to Xero.
- EROAD telematics and weighbridge-docket OCR shown as live integrations
- AI-assist that turns a day's field notes into invoice lines a person approves
- Plant-and-machinery visual identity built to feel like real operations software
BuildLedger Five systems, one cost code.
An integrated finance and estimating platform for a New Zealand quantity-surveying firm, joining CostX, Xero, EzzyBills, ApprovalMax and Microsoft 365 into one interface.
- Budget-vs-actual reconciliation line by line, on the cost code both systems share
- Exception rules that surface what needs a person, with source attribution on every figure
- A read-only AI assistant and the NZ statutory layer — payment-claim clocks included
Orchard Designing out six weeks of spreadsheets.
A targets platform being designed for an apple grower: set next season's crop (T/ha) and fruit-size targets per block, then track target against actual from orchard and packhouse exports.
- Designed so every derived number is traceable — the API returns inputs and a formula id with each output
- All formulas in one module; weights and areas as decimals, never floats
- Idempotent CSV import, multi-season and multi-orchard from day one, ready for AI forecasting
Aion The camera sees; the cloud counts.
A family of edge-camera analytics products. In OpsLens, cameras on the factory floor publish on-device analytics events; the cloud turns them into OEE, machine state, downtime and production counts.
- No raw video leaves the site — only events, validated against a versioned JSON Schema contract
- One shared, business-agnostic camera package; each product owns its own backend
- Multi-tenant from the first customer, deployed across Fly.io, Koyeb and AWS
Cairn An agent that knows what it did not read.
A file-driven agent harness for licensed immigration advisers in New Zealand and Australia: case consistency checks, pre-submission review and drafted replies to official letters.
- Extracted facts, conflicts and checklist verdicts are validated by code before anything is shown
- Agent runs write back to memory; a weekly cleanup consolidates it only with human approval
- Every model call names a role, so vendor, model and rate limit change in config, not code
llm-gateway One endpoint for every project.
A local gateway that gives all my projects a single OpenAI- and Messages-API-compatible endpoint, routed to named profiles — so applications use the official SDKs and never hard-code a vendor.
- JSON mode, streaming, client tools and image input behind the same URL
- Per-profile daily request and token limits, answered with a standard quota error
- A
doctorcommand that checks setup without spending quota; runs privately on a VM over an SSH tunnel
AdForge One spec, three kinds of ad.
A spec-driven engine that turns a product image and a brief into 10–30 second video ads and static image ads for every platform — and a raw screen recording into a polished product demo.
- A single typed spec is the only contract between planning, rendering and review stages
- Planner and critic agents, chosen by measured evaluation rather than default model
- Demos get auto-zoom, step callouts, captions and device frames; one output per platform size
§ 03 — Principles
How I work
Code decides; the model drafts.
A model can propose a fact, a verdict or an invoice line. Code checks it against the contract before anyone sees it.
Coverage is a first-class output.
A gap is reported as a gap. Silence is never allowed to look like a clean result.
Find the assumption everything rests on.
Before building, I name the one join or key that must hold — and test it on real exports first.
Evidence before “done”.
Tests that fail first, builds that exit clean, screenshots of the real screen. Reviewing evidence is faster than trusting it.
§ 04 — Contact
Have a process that a model could help with — but must not get wrong?
arthurgao@outlook.com github.com/ArthurGao
Based in Auckland, New Zealand