Research for AI systems that have to work in the real world.

Governed agents

High-impact agent actions remain permissioned, reviewable, interruptible, and explainable.

CloseGate demonstrates action tiers, segregation of duties, mandatory approval, materiality routing, and hash-chained audit. Regulus adds identity, purpose, privacy, residency, model-risk tiers, kill switches, and evidence export. Neither system has a published external certification or third-party deployment. At Aveni, I was part of the team in the first FCA Supercharged Sandbox cohort, working on AI for wholesale and compliance innovation.

Durable and coordinated agent systems

Agent work survives crashes, concurrent writers, and the distance between a decision and its side effect.

fast-langgraph explores durable state and checkpointing. Before the Pull Request studies coordination before code review; in its measured setup, duplicate or conflicting rework fell from 78% to 0%, useful throughput more than tripled, and replicas converged without silently dropped writes.

Exact evaluation for generated code

A fast generated kernel is useful only after the team can establish that it is correct.

gpuemu uses operator-aware test generation and reproducible failure artefacts for CUDA and Triton kernels. In the measured corpus, its seeded oracle caught 10 of 10 injected defects while 16 of 16 correct controls stayed clean across five GPU classes. That result does not estimate the defect rate of deployed LLMs.

Local inference and physical AI

Keep model placement, physical constraints, review, execution, and recovery explicit from cloud plan to edge action.

At Orangewood Labs I built RobotGPT, which reduced collaborative-robot programming time by approximately 10× in that specific context, and worked on computer-vision and cloud-to-edge systems. WareMax is a separate deterministic warehouse benchmark for task allocation, reward design, causal delay attribution, and byte-identical replay. Cognisoc explores local inference across multiple language runtimes.

Cloud and Day-2 AI operations

The deployment is the beginning: ownership, observability, budgets, secrets, retries, evaluation, rollback, and stop conditions make it operable.

Recent delivery includes agents on Azure AI Foundry and AKS/Kubernetes, custom sandboxing, and MCP in .NET. Earlier work includes Python and AWS document extraction, validation, routing, and LLM workflows. Customer architectures and scale remain confidential.

Blockchain infrastructure and safety boundaries

Make signing policy, cross-chain state, compiler semantics, and agent authority explicit before money or keys move.

Cryptuon projects cover validator signing policy, atomic multi-rollup actions, commit-reveal timing, EVM and Solana compiler boundaries, and paper-first agent-operated markets. They are public design and implementation artefacts; they do not establish audits, mainnet adoption, returns, or production scale.