AI Engineer
that ships.

George

AboutGWork

A short note

AI Engineer · Cloud & Systems background · Melbourne

I build AI products that ship to real users.

I'm George. I design, build and deploy AI applications end to end: agent pipelines, retrieval, evaluation, and the unglamorous infrastructure that keeps them running.

By day I run technology for a 40+ store Australian retailer, where the systems I build carry real transactions. After hours I build under GEOSTACK, mostly compliance and research products where a wrong answer has consequences. That's why every figure in my apps goes through deterministic tools, never through the model.

I'm looking for a founding or forward-deployed AI engineering role at a startup that wants someone who ships.

— George

See the workGitHub ↗
  • LangGraph
  • NVIDIA NIM
  • Qdrant
  • Supabase
  • FastAPI
  • Next.js
  • Ollama
  • MCP
  • Railway
  • Vercel

Selected work

Five projects, running.

01 · Live · Compliance copilot · NDIS

CareCert

An AI-first compliance copilot for NDIS providers facing mandatory registration. Providers manage registers, incidents, workers, evidence and audit packs, with a chat copilot grounded in official guidance. A deterministic rules spine decides compliance; the model only explains.

  • Next.js 14
  • FastAPI
  • LangGraph
  • Qdrant
  • nv-embed-v1
  • Supabase RLS
  • Kubernetes
Visit ↗

Worth asking me about

  • 145 backend tests, made hermetic with a gateway stub so the suite runs in 35 s with no live calls
  • Real RAG on NVIDIA embeddings (4096-dim) over official guidance
  • Supabase auth on ES256/JWKS, org provisioning, row-level security in the Sydney region
  • Mapped which Australian registers have real APIs (ABN Lookup, ABLIS) and which don't
ndiscert.vercel.app
Illustrative demoA compliance checklist ticking off six items beside a chat where the copilot answers a question with three cited sources.

02 · v1 → v7 · Research agent · RAG

Research Conductor (AskMyDocs)

Began as document Q&A and grew over seven versions into an autonomous research pipeline. A LangGraph orchestrator scrapes several academic databases in parallel, routes across models, then runs a Writer↔Critic loop until the draft passes a rubric.

  • LangGraph StateGraph
  • Multi-model routing
  • Qdrant
  • Streaming

Worth asking me about

  • Self-correcting Writer↔Critic loop with a Pivot agent when the evidence runs dry
  • Two-layer guardrails and rubric-based confidence scoring
  • Bibliography generated programmatically, never by the model
  • Memory agent degrades gracefully when the vector store is down
research-conductor
Illustrative demoA state graph: an orchestrator fans out to three scrapers, a synthesiser, then a writer and critic loop until a rubric score passes and output is produced.

03 · Built · Multi-agent · Finance

FinSight

A multi-agent investment research platform. Nine specialist agents with typed Pydantic outputs run in parallel and stream results to the UI over WebSockets.

  • NVIDIA NIM
  • Pydantic
  • asyncio
  • WebSockets

Worth asking me about

  • Async rate limiter: semaphore at 2 concurrent calls with jittered backoff, so it runs on a free inference tier
  • Every number comes from a tool call; the model never generates figures
finsight
Illustrative demoNine agent lanes, two active at a time under a semaphore, each finishing with a JSON result before a summary card assembles.

04 · Built · Fine-tuning · Local

QLoRA JSON extractor

A small open model fine-tuned with QLoRA on a consumer RTX 3050 under WSL2, so it returns reliable structured JSON without a long system prompt. I measured extraction accuracy before and after tuning on the same test set.

  • QLoRA
  • PEFT
  • WSL2
  • Ollama

Worth asking me about

  • Fitting training into 4 GB of VRAM
  • When fine-tuning beats prompting, and when it doesn't (I kept RAG for CareCert)
qlora-extractor
Illustrative demoA terminal printing training logs beside a loss curve drawing itself, then raw text becoming validated JSON.

05 · In use · Evals · Observability

Shared eval & monitoring layer

One evaluation layer reused across CareCert, Carbonly and Research Conductor. It logs every AI call's output, latency and cost, and scores answers against a hand-written set of known-correct cases.

  • Golden sets
  • Cost tracking
  • Latency
  • Sentry

Worth asking me about

  • Writing golden sets for regulatory answers
  • Catching regressions before a prompt change ships
eval-layer
Illustrative demoA streaming log table with pass and fail pills, a pass-rate sparkline, and a regression caught on a prompt deploy then rolled back.

Backdrop: Gibson Hazard, “Camera Roll”

More builds

In build · Specs → tickets

FieldCert

Rules engine and evidence verification for energy-efficiency certificate schemes (VEU, ESS), with an immutable evidence ledger and AI fraud checks.

Agents · Orchestrator + 4

RetainAU

Agents that track construction retentions and bank guarantees and draft release letters when they fall due.

Full stack · Scraping

PriceWatch

Competitor price monitoring across ten Australian retailers: LLM-assisted scraping into FastAPI and Supabase, with a Next.js dashboard.

Live · exilnk.vercel.app

Exilink

Work-defence verification: candidates defend decisions in their own project work, as proof of understanding.

In build · RL + QLoRA

HVAC digital twin

A thermal model of a building, a PPO agent that optimises HVAC setpoints, and a fine-tuned LLM that explains the agent's choices. Mechanical engineering meets RL.

Local-first · Fine-tuning

Email persona model

A fully local model that drafts replies in my own writing style. It drafts only and never sends; private data stays on my machine.

Production

Running in production

Systems I built and run for a 40+ store Australian specialty retailer.

2000 →

Order history backfilled

Commerce data pipeline

BigCommerce to Supabase sync on a nightly GitHub Actions schedule, with full historical backfill and row-level security on every table.

~95%

AI participation in support chats

AI support monitoring

KPI reporting on the customer-support AI agent through an Intercom MCP integration. Self-resolution runs at roughly 75–89% depending on the period.

26

Stores synced

Store data middleware

Serverless sync that treats the reviews platform as the source of truth for store hours and details, plus llms.txt work so AI assistants can find the business.

40+

Stores supported

Systems & security

POS and ERP (LS Retail on Dynamics NAV), Power BI reporting, an Intune-managed iPad kiosk rollout, and identity hardening across Microsoft 365.

What I work with

Skills

Only what I'll defend in an interview.

AI systems

LangGraph,multi-agent orchestration,RAG,Qdrant,NVIDIA NIM,Ollama,QLoRA,evals,MCP servers

Production and portfolio

Product engineering

Python / FastAPI,TypeScript / Next.js,Supabase,Expo,Stripe,Resend,Sentry

My standard stack

Infrastructure

Docker,GitHub Actions,Vercel,Railway,AWS & Azure fundamentals,Kubernetes,Terraform

Kubernetes and Terraform at project scale

Business systems

BigCommerce,LS Retail / Dynamics NAV,Power BI,Microsoft 365,Intune,Freshdesk

Day to day at HSW

AI-assisted dev

Claude Code with custom skills for writing review, coding guidelines, academic research and analytics reporting

Daily workflow

Credentials

Master of Research Studies

In progress · Torrens University Australia

Thesis: AI-based vulnerability scanning in DevSecOps pipelines in Australian healthcare IT.

Building with Claude

Completed · Anthropic Academy

Claude Code 101, Introduction to Subagents, Introduction to Agent Skills, Building with the Claude API.

Mechanical engineering

Foundation

Systems thinking and physical modelling. It shows up in the HVAC digital twin.

Open to founding and forward-deployed AI roles

Let's build something real.

GitHub ↗Email