Expertise

Automation & Applied AI

Automation for expensive manual work, decided step by step: full automation, a human in the loop, or no software at all. When a model is part of the answer, it comes with evals, guardrails, and a measured cost.

Process automationHuman-in-the-loop designEval engineeringModel selectionCost optimizationDocument processing

Industries

What it covers

Process mapping and baseline measurement before anything is built
Clear calls on full automation, human-in-the-loop, or leave manual
Eval suites that make model and prompt decisions measurable
Model selection and routing on measured cost, latency, and accuracy
Deterministic guardrails around any stochastic component
Document extraction and classification pipelines

Typical work

Back-office and operations workflows where staff time is the bottleneck
Document extraction and classification for intake and research work
Eval sets that make model and prompt decisions defensible
Cost and latency reviews for model-backed features already in production

How I think about it

Consequential automation is an engineering problem before it's an AI problem. The first question is whether a step should be automated at all. When a model earns a place, I treat it like any other dependency: measured, tested against an eval suite, and replaced when a cheaper one does the job.

Other expertise

Agentic Engineering

Coding agents as part of the engineering system. Harnesses, guardrails, quality gates, and telemetry.

Platform & Infrastructure

AWS and GCP platforms defined in code. Kubernetes, GitOps delivery, observability, and cost control.

Security & Compliance

Identity, least privilege, audit trails, and data boundaries designed in. Architecture that holds up under HIPAA, SOC 2, and ISO 27001 review.

Data Engineering

Ingestion, enrichment, and event pipelines. Data modeled and accessible for products, analysis, and automation.

Product Engineering

Web, mobile, and backend products from first version through launch and iteration. Full stack, from data model to interface.