Built, operated, and secured production software and infrastructure. Researching agent memory, evaluation, and the limits on what model-written code can reach.

I work in the open at alignment.farm and in alignment-farm and stack-research.

• construct: research lab for governed agent memory, in Python. The harness chooses the context a model receives and scores the result against an external check. Spoofing an unauthenticated memory channel changed answers; authenticating that channel stopped the spoof. Pruning cold state and rematerializing it cut hot-context tokens by about 57% at matched answer quality. On a continuity test, governed memory passed 24 of 24 relevant cases against 6 of 24 for canonical state alone.
• construct-2: studies of how an agent accumulates experience across sessions, and where it should live: explicit memory, lessons, executable tools, weights, or runtime rules. The clearest result is that learning a pattern, keeping it useful, and repaying its cost are separate. On a directly addressable collection, cheaper explicit access matched learned retrieval.
• substrate: local-first shared rooms, in Go. People, agents, and tools take turns in one conversation. The room is an append-only directory of Markdown, and every interface uses the same turn rules.
• materialize: builds an isolated workspace from a read manifest. Only declared paths are copied, a denylisted path aborts the build, and a hash audit records exactly what was present. Used to keep a red team, and separately a builder, inside a declared file surface.

• The Unaskable Question: mapping the negative space of language models (field-note, in progress)

Engineering consultant and former CTO. Experience in building systems, owning production infrastructure, security, architecture, technical leadership, and self-directed experimental work.

• Agent memory and evaluation
• Applied model systems
• Cloud infrastructure and legacy modernization

Languages: Python, TypeScript, Go, Ruby, Bash.

Applied models: LLM integration, retrieval-augmented generation (RAG), structured generation, context engineering, agent and tool workflows, model harnesses, persistent agent memory, PyTorch, Ray, LoRA fine-tuning, reinforcement learning from verifiable rewards, evaluation harnesses, experiment tracking, Bedrock, SageMaker, MLX, vLLM-Metal, llama.cpp.

Data and infrastructure: Spark, Parquet, PostgreSQL (pgvector), S3 Vectors, SQL, NoSQL, key-value, and vector data modeling, streaming pipelines, fault isolation, Docker, Kubernetes, AWS (CDK v2, CloudFormation, Terraform, CodePipeline, EventBridge), distributed tracing.

Security: Application and cloud security, threat modeling, penetration testing, sandboxed execution (gVisor, Firecracker), tool permissions, auditable workspaces, zero trust, secrets management, SOC 2, HITRUST.

Architecture: System and API design, event-driven and service-oriented systems, incremental legacy migration, dependency decomposition, technical risk assessment, technical writing.

Certified: AWS Certified Generative AI Developer – Professional (Issued December 2025)