Selected work

Built, not just recommended.

Each write-up covers the problem, what we built, the stack, and what it shows about how we work. Client names are withheld; demos and walkthroughs are available on request.

Healthcare · Agentic AI · Demo

Prior-authorization AI intake

Prior authorization arrives by fax, PDF, and payer portal, and nurses spend hours re-keying it before any clinical judgment happens. This demo shows what an AI-assisted intake looks like end to end — with clinicians firmly in control.

Claude
Document extraction
Workflow UI
Dashboards

What we built: Multi-channel intake pipeline for fax, PDF, and portal submissions

Claude-based extraction of clinical and administrative fields from unstructured documents

Nurse review worklist where reviewers can accept or override every AI recommendation

Operations dashboard for volume, turnaround, and queue health

A CMO-facing slide deck explaining value, controls, and rollout
Why it matters: It demonstrates human-in-the-loop design for a regulated workflow: AI removes the clerical load, while accountability and final decisions stay with licensed clinicians.
Status: Working demo. Walkthrough available on request.

Health plan · Strategy · Engagement kit

AI discovery workshop

Leadership teams often agree they "need AI" without agreeing on where, why, or how to govern it. This anonymized workshop kit, built for a health plan, turns that ambition into shared priorities.

Executive facilitation
Use-case prioritization
Roadmapping
Governance

What we built: A facilitated discovery workshop deck and structure that moves executives from goals and pain points through candidate use cases, data and risk considerations, and a prioritized, sequenced roadmap.
Why it matters: It shows the boardroom-to-technical translation at the heart of the practice: framing AI in business terms leaders can decide on, while keeping every recommendation buildable.
Status: Anonymized sample available on request.

Local-first ML · LLM · Python

Local-first document tailoring pipeline

A personal-productivity system that tailors structured Word documents to a target description — a compact example of combining small local models with a frontier LLM, and keeping data on the machine where possible.

Python
Hugging Face
Embeddings
Zero-shot classification
Claude API

What we built: Semantic matching with local embeddings (BAAI bge-small-en-v1.5)

Zero-shot industry classification (DistilBART-MNLI) to decide which experience to foreground

Canonical title matching against the O*NET alternate-titles dataset

LLM-assisted rewriting through the Claude API

Templated Word output with docxtpl and Jinja2
Why it matters: It's a hands-on pattern clients ask about: use cheap, private, local models for the steps they handle well, and reserve the frontier model for the steps that need it.
Status: In active use and development.

Cloud architecture · Security

This website

A small, deliberately boring reference architecture: a static site with no third-party scripts on Azure Static Web Apps, with a serverless contact form that sends through Microsoft 365.

Azure Static Web Apps
Azure Functions
Microsoft Graph
Cloudflare DNS

What we built: Hand-written, accessible HTML and CSS; strict security headers (CSP, HSTS, frame and referrer controls); a contact form with server-side validation and spam traps; and automatic, repeatable deployment from GitHub.
Why it matters: Secure-by-default, inexpensive to run, and easy to audit — the same instincts we bring to client architecture.

Have a workflow worth fixing?

Tell us about it. We'll reply personally with a clear next step.

Start a conversation