01 — AI training & evaluation

Engineering rigour, applied to model evaluation

A year building what frontier labs are short of: hard, unambiguous, verifiable engineering problems with objective acceptance criteria.

Frontier AI
Turing
Jul 2024 — Aug 2025

CUDA-EVAL — GPU-programming evaluations for frontier models

A year on CUDA-EVAL, designing and reviewing GPU-programming evaluation tasks used to train and benchmark advanced language models. The hard part is not writing CUDA — it is writing a problem that has exactly one defensible answer, and a rubric that can tell a correct solution from a merely convincing one.

  • Authored evaluation tasks with explicit inputs, assumptions, constraints and objective acceptance criteria, so grading is reproducible rather than a matter of opinion.
  • Produced reference solutions and validation checks, confirming correctness analytically and by execution rather than by inspection.
  • Peer-reviewed other engineers' submissions for technical correctness and completeness, feeding consistent quality standards back into the task set.
  • Applied the same discipline used in safety-relevant engineering review, where a plausible-looking wrong answer is the expensive failure mode.
Directly relevant to simulation-led AI training work: problem definition, reference modelling, analytical validation and peer quality assessment — the same loop, whether the domain is CUDA kernels or electrical machines.
CUDAEvaluation designReference solutionsPeer reviewTechnical writing
Applied AI
PathPartner Technologies
2022 — Aug 2024

Language models on automotive silicon

The other side of the same coin — not evaluating models, but making one survive contact with an embedded target.

  • Deployed a 7-billion-parameter LLM for in-cabin voice command understanding, tuned to the SoC's memory and compute envelope.
  • Built the speech-to-text front end, benchmarking Google SpeechRecognition, Azure Cognitive Services and IBM Watson against on-device performance.
  • Validated across real driving conditions — cabin noise, speech variation, degraded input — because lab-only accuracy is not a result.
Llama 2 (7B)NLPSpeech-to-textEdge inference