Independent AI systems testing and deployment engineering.
I test model-serving stacks on real hardware, publish the conditions and raw data, and keep the failures, reversals and corrections visible alongside the wins.
My background is hands-on data-centre deployment — B200/B300-class GPU systems, fibre, commissioning and maintenance — and I bring the same discipline to the multi-node model-serving work in my own lab.
Registry counts are generated from model-serving-minefield at fb3acb1, 2026-11-08 — not typed into this page by hand.
Available for
Independent testing, diagnosis and deployment work where the answer needs evidence rather than a guess.
Independent claim verification
Testing a vendor or community claim under controlled conditions, with the raw data published whichever way it lands.
Evidence: HumanEval+ replication · identity-prefix study
Serving and deployment diagnosis
Working out why a model looks broken when the model is fine — often a template, parser, history or serving-stack problem rather than a weights problem.
Evidence: multi-turn reasoning-stripping · model-serving-minefield
Quantization verification
Establishing whether a third-party quant actually holds: does it build and serve, what does it cost in memory, and does quality survive against the control build?
Evidence: quant-floor study · quant label ≠ kernel path
Long-duration production soaks
Running a model inside a real agent pipeline for as long as the question needs, and reporting drift, incidents and integrity behaviour rather than a single throughput number.
Evidence: 12-hour production soak
Multi-node and tensor-parallel deployment
Bringing up models across networked nodes: fabric, topology, serving configuration and the failure modes that only appear once a model spans more than one box.
Evidence: Hy3 on dual DGX Spark · MiniMax-M3 on dual DGX Spark
Eval-harness and measurement auditing
Auditing a benchmark before trusting it: arm-level handicaps, scorer normalisation, contamination screens and harnesses that quietly score a correct answer as wrong.
Evidence: HumanEval+ replication · Minefield evaluation checks
A useful first pass
If you are debugging an OpenAI-compatible endpoint, start with minefield-doctor and send me the report together with the number or behaviour that looks wrong. That gives us evidence before we start changing things.
Contact
Email or DM me on X. Include the stack you are using and the number or behaviour that looks wrong; that usually makes the first reply useful. If the finding belongs in the registry, open a model-serving-minefield issue so the evidence and credit stay public.