Independent consulting practice
I help engineering teams move AI work from prototype to production. Most of what I do sits at the intersection of cloud platforms, agent design, and the operational reality of running these systems once they leave the demo.
TL;DR
AI projects framed around a decision someone makes or a workflow someone uses.
Agent systems with explicit tool boundaries, real evaluation suites, and human oversight at the points that matter.
Cloud and IoT delivery measured by what it costs to run and how often it breaks, not by the polish of the demo.
What I do
Most engagements span more than one. The interesting problems usually live at the boundaries between them.
Moving from interesting model output to a capability that fits a real product. The work is selecting the right use case, plumbing in the data, and designing the operating model around the AI itself.
Agent loops fail in predictable ways: unbounded tool use, drifting context, missing evals. Designing them properly is most of the work; running them in production is the rest.
The infrastructure that quietly determines whether a product survives its first thousand users. APIs, event flows, deployment automation, security boundaries, and a FinOps model that holds up at scale.
Connecting hardware telemetry to dashboards a human actually looks at. Includes the mobile and web work that closes the operational loop between the field and the office.
Turning a strategic direction into a sequence of engineering decisions: discovery, target state, migration paths, and the governance to keep delivery on track without drowning the team in process.
How a system holds together
A request lands, gets validated, picks up the data that grounds the answer, runs through a model with scoped tools, and writes an audit record that holds up later. The architecture is mostly about what happens between those steps.
Request -> Queue -> Retrieval -> Model -> Tooled action -> Audit
Token budgets, prompt-injection defenses, request quotas, fallback to cheaper models, and the regression suites that catch a silent model swap before customers do.
Approval queues, draft-with-edit surfaces, and override-and-re-prompt flows. The thin layer that turns mostly-right output into something an operator is willing to sign.
$ inspect run
ingest: ok
retrieval: 12 chunks
model: ok (1.2s)
evals: passing
audit: signedWhat I help build
AI tied to a workflow, a data source, and a decision someone has to make. With the eval and feedback machinery to keep it honest once it leaves the prototype.
The loop, the tools, the memory model, the permissions, the audit trail, the eval suite, and the human checkpoints. Each one is a design decision worth taking seriously.
APIs, events, data pipelines, identity, observability, and the deployment automation that lets a small team operate a serious system without burning out.
Telemetry pipelines, fleet operations, field workflows on mobile, edge and cloud boundaries, and alerting that respects an on-call schedule.
Legacy decomposition, migration sequencing, integration strategy, debt reduction, and team enablement that outlasts the engagement itself.
How engagements run
Clarify the outcome, the constraints, the integration surface, and what should not be built. One or two weeks, depending on scope.
Architecture decisions, ownership, security boundaries, delivery practices, cost model, and the signals that say it is working.
Prototype the critical paths, review the implementation, coach the team, and turn open questions into shipped code.
Twenty years across banking, government, fintech, and platform engineering. The practice exists because the systems worth building usually pull from more than one of those worlds.
Cloud modernization in banking and fintech: on-premises to AWS migration, platform tooling, CI/CD, and FinOps practice.
Engineering across Java, Kotlin, Go, Python, Node.js, and Android, plus the data and automation work that surrounds them.
Proofs of concept, stakeholder alignment, mentoring, cloud migration, and the discovery work that decides what should be built.
Read the full CVGet in touch
Send the goal, the current stack, the constraints, and the timeline. I will reply with a concrete next step or, if it is not the right fit, a suggestion of who else to talk to.