Work · architecture · data · operations

Production AI fails at the seams.

I build and review the systems at those seams: where models meet state, tools meet authority, data meets context, and a promising demo meets operating reality.

Three ways to work together

Choose the intervention that matches the problem.

Architecture review

A focused, evidence-backed assessment of what an existing or proposed agent actually does, where it can fail, and what must change before the next production decision.

View the review process → 02

Reference build

One difficult production seam built to a usable standard—memory, evaluation, data foundations, execution, or control—then handed over with runbooks and decision records.

See the reference builds → 03

Embedded enablement

Work alongside the team until they can run, extend, and govern the system themselves. Capability transfer is part of the deliverable. Offered only after a review verdict, not as a parallel entry.

Follows the review →

Selected proof

Built around production seams.

All reference builds →

Agentic AI

Durable agent execution

Checkpoints, cancellation, idempotent effects, and replayable failure evidence for work that outlives a process.

Read the case study →

Agentic AI

Agent memory over a lakehouse

A public typed, scoped, bi-temporal memory implementation, with 59 tests and six scenarios rerun.

Inspect the verified case study →

Agentic AI

Verified-capability routing

A public, tested implementation of requirement-driven selection, trust checks, capability decay, and reconstructable routing evidence.

Inspect the verified case study →

Reference-build archive

Eleven implementations you can inspect.

Agent systems6 builds

Verified-capability routing for multi-agent systemsDynamic agent selection based on verified capabilities and trust boundaries. Durable agent execution with checkpoints and cancellationResumable runs, idempotent effects, explicit cancellation, and replayable evidence. Agent memory over a lakehousePublic implementation; 59 tests and six scenarios rerun on 3 September 2026. Evaluation and observability harness for agent systemsTraces, replay fixtures, eval gates, drift signals, and cost attribution. Agent privacy as a data-flow architectureContext minimization, contract-bound egress, scoped retention, and a privacy decision ledger. Exception queue before happy-path automationStructured human handoff, policy-driven escalation, recovery states, and an unresolved-work ledger.

Data foundations2 builds

Enterprise RAG document extraction pipelineLayered extraction, coverage audit, and human-in-the-loop exceptions. Sutra Feed GuardDeterministic drift classification and governed change before production data is touched.

Platform and control3 builds

Agent sandbox sized to blast radiusIsolation, capability-granted mounts, resource limits, egress controls, and audit trails. LLM gateway as the control planePolicy-controlled redaction, eligibility, quality, residency, spend, routing, and fallback. Policy-as-code for enterprise agent governanceVersioned rules, approval gates, deny precedence, runtime enforcement, and audit evidence.

Current portfolio

A growing set of production seams.

Eleven reference builds currently cover memory, evaluation, privacy, routing, execution, data foundations, exception handling, sandboxing, model control, and governance. New use cases can be added as the work evolves.

11 builds so far

How I work

The team should be able to run the system after I leave.

Top-downThe shape of the whole is decided before any component is developed in depth.

Verification-orientedGenerated content, vendor claims, and first drafts are checked rather than trusted.

TransferableRunbooks, architecture decisions, and progressive training are part of the work.

Have a production question?

Tell me what you are building, where it is stuck, and what “done” means.

Request a Production AI Architecture Review More about how I work →