Six examples across consumer machine learning, robotics, generated-code evaluation, regulated banking AI, blockchain venture building, and Azure agent delivery.
Consumer ML at scale
Built or led work across multilingual NLP, recommendations, personalisation, trust and safety, computer vision, datasets, offline evaluation, inference, A/B tests, and live-signal iteration.
Scope: This was team delivery at company scale; the company and product outcomes were collective.
Physical AI
At Orangewood Labs, built RobotGPT and worked across computer vision, cloud-to-edge systems, and multidisciplinary ML, backend, edge, and product engineering.
Scope: The approximately 10× reduction was measured in that collaborative-robot programming context.
Generated-code correctness
The operator-aware oracle caught all ten injected CUDA/Triton defects while all 16 correct controls stayed clean across five GPU classes in the measured corpus.
Scope: The study used a controlled seeded corpus; it did not estimate defect prevalence in deployed LLM-generated code.
Regulated banking AI
Architected enterprise banking AI including Agent Assure, conduct-risk workflows, evidence generation, human review, escalation, evaluation, versioning, and release controls. Participated as part of the Aveni team in the first FCA Supercharged Sandbox cohort.
Scope: The FCA lists Aveni among the accepted wholesale and compliance firms. Customer details, architecture, volumes, and security outcomes remain confidential.
Blockchain venture building
Boom Protocol secured more than $2.5M in investor commitments and commercial letters of intent while developing blockchain-infrastructure concepts.
Scope: The record contains commitments and letters of intent, not a completed raise, contracted revenue, or production deployment.
Azure agent delivery
Recent Neul Labs delivery includes Azure AI Foundry, AKS/Kubernetes, custom sandboxing, and MCP in .NET.
Scope: Customer identity and usage scale are confidential; no public certification or security-effectiveness study is available.
Start with the outcome, the decisions at stake, and the constraints that cannot be wished away.