Field notes from production
Most of the work is not the model. It is everything that keeps the model alive: vector search, retrieval, serving, latency budgets, the pipeline that survives Monday morning traffic.
The difference
Any engineer can ship a model.Few keep one aliveunder twenty million requests a day.
Requests a day served by a serving architecture I built for ImagineArt: 30+ models behind FastAPI, Docker and Kubernetes.
Vyro
Uptime held at peak load across that same fleet.
Vyro
Search relevance improvement from replacing keyword lookup with vector plus hybrid retrieval.
Stealth startup
Latency cut on real-time mobile body tracking by re-tuning the MediaPipe to Unity path.
Add Life
Faster time-to-ship for new generative features, via a reusable Stable Diffusion hosting SDK.
Vyro
Fewer production errors after rebuilding CI to 97% coverage.
Vyro
A session-based job copilot that analyses fit, tailors a resume and drafts outreach, remembering across threads.
A bespoke inference architecture for ImagineArt, then the second-largest AI art generator in the US.
Name a dish; RAG returns the recipe with every ingredient matched to a real product, added straight to the basket.
Serverless fine-tuning on Runpod and AWS cut a personalised avatar model from hours of turnaround to 15 minutes.
Vector search over hundreds of thousands of products lifted relevance by 35–45%.
Production systems without a write-up, open-source tools, and the research pipelines behind them. Same record, less narrative.
Open the full indexJul 2024–Jan 2026
Stealth Startup
Sydney, Australia
Rebuilt product discovery for an Australian grocery platform around vector retrieval.
Feb 2025–Jul 2025
Add Life Technologies
Adelaide, Australia
Made real-time body tracking usable on mid-range phones, cutting latency by 40%.
Aug 2022–Jan 2024
Vyro
Wyoming, United States
Owned the inference layer behind a top-two US AI art generator.
5.2 · Model · 6 frameworks
Training and evaluation, mostly PyTorch, mostly on data that arrived messy.
The University of Adelaide
Work from this period
National University of Computer and Emerging Sciences, FAST-NUCES
Work from this period