Services
Four things, done well.
I take on a small number of engagements at a time. Every capability below is drawn from work that’s either live in production or built end-to-end with the same standard.
01
Production web
& commerce
Multi-vendor commerce, payments, fulfillment, role-scoped admin.
The systems that handle money, ship product, and answer to auditors. I’ve built one end-to-end and I know where the edges are: idempotency on every payment path, server-side price re-resolution so the client can’t tilt the math, an immutable ledger when balances need to reconcile, careful permission scoping so a vendor can’t see another vendor.
Typical scope: a marketplace from scratch, a payments overhaul, adding multi-tenant data separation to an app that grew past single-tenant.
Built with this
Who’s Brew — multi-vendor coffee marketplace →02
Applied LLM
features
LLM features that ship and that you can afford to run.
Natural-language search, personalized recommendations, content generation, structured extraction. Choosing the model and the inference provider is the small part. The interesting part is designing the prompt and the data flow so the answer quality is consistent, the cost per request is bounded, and the feature degrades gracefully when the model is wrong — which it will be.
Typical scope: adding a real LLM feature to an existing product, replacing a clunky search experience, building a recommendation engine that learns from real user behavior.
03
Computer
vision
Local-first CV pipelines for documents and workflow automation.
Capture, clean up, vectorize, hand off. Hybrid pipelines that combine classical OpenCV stages with neural models where they earn their place. Hardware acceleration through ONNX DirectML or OpenVINO when the workload justifies it. Local processing when the data shouldn’t leave the user’s machine.
Typical scope: digitizing physical documents, perspective correction and OCR pipelines, vector output for downstream editing, integration with existing storage and notebook systems.
Built with this
OneNote Whiteboard Scanner — local CV plus Microsoft Graph →04
ML model dev
& deployment
Train a model, export it, serve it as something small.
Training in PyTorch or TensorFlow, exporting to ONNX, serving on CPU through a thin REST API that deploys anywhere — App Runner, ECS, a $5 VPS, or a customer’s own infrastructure. The discipline is in keeping the inference footprint small enough that the model is actually usable in production, not just impressive on a laptop.
Typical scope: prototyping a custom model for a specific problem, taking a research-grade model into a deployable service, replacing a third-party API with an owned model when cost or privacy demands it.
Built with this
AI Chess Engine — AlphaZero-style ResNet + MCTS, ONNX REST →Get started
Most engagements start with a 30-minute call.
Tell me the shape of the problem. I’ll tell you whether it’s something I can help with, what an engagement would look like, and what it would cost.
Book a call