Work
Client work, and the systems I build and run
Client case studies first, then the AI systems I've designed, built, and run myself: agents, a multi-tenant RAG platform, native AI assistants, local LLMs, and AI infrastructure. All working code, most of it in production.
Case Study
Automating a matchmaking firm's intake and consultation pipeline
A boutique matchmaking business was collecting client intake by hand and re-keying it, along with notes from every consultation call, into their matchmaking CRM. I built the pipeline that removed the manual step.
What I built
- A custom Angular intake form on a Spring Boot backend, with draft autosave, field-level revision history, and validation
- On submit, an asynchronous push of the intake data into their matchmaking CRM (SmartMatchApp) over its REST API, with an email notification linking straight to the newly created client profile
- A Zoom integration that captures AI-generated consultation summaries via webhooks and a Server-to-Server OAuth app, then parses and feeds them into the same CRM
- Business email migration, hosting, domain cutover, and analytics and privacy setup to take it all live
The outcome
Client intake and consultation notes now flow into the matchmaking CRM automatically instead of being typed in twice by hand. Launched in production in April 2026.
Case studies
Real findings from paid code reviews, free QA passes, and informal looks at founders' apps, all anonymous. A full AI App Review goes further than any of the outside passes.
A CRM in closed beta
Three crashes, a record cap that a restore could walk straight past, and quotes that saved with any price you typed.
Read the case studyA beverage community app
A "private" setting that didn't always hold, and form rules the server never enforced.
Read the case studyRecords that stopped at 5,000
A paid code review found an AI-added limit that silently cut exports, reports, and search off at about 5,000 records.
Read the case studyCustom fields the database couldn't search
Every custom field stored in one block of JSON, so pages got slower with every record a customer added.
Read the case studyA secret silently replaced
Saving a secret with a name already in use overwrote the old value and said "Saved." The team fixed it.
Read the case studyA live feed showing hardcoded data
When the real data source failed, the site quietly showed bids written into the code instead.
Read the case studyBilling pages that disagreed
A low-balance warning on a new account, a minimum deposit the site rejected, and costs that changed by page.
Read the case studyA plugin update that wouldn't stop running
An automatic update looped forever on a store seven hours from UTC. I traced it to two bugs and reported them to the vendor.
Read the case studySystems I build and run
Personal AI Agent Infrastructure
A self-hosted MCP server that turns Claude into an autonomous operations assistant: 59 tools across 12 functional domains, running in production every day.
Remote MCP Servers for AI Connectors
Two live products expose their own remote MCP server - 73 tools and 21 tools - that Claude and ChatGPT connect to over the web, secured by an OAuth server I wrote so each connector is scoped to a single tenant.
Multi-Tenant RAG Chatbot
A full product: passwordless auth, per-org isolation, local ONNX embeddings, grounded answers with citations, and per-message cost tracking.
Klass CRM - AI Assistant
A native macOS/iOS CRM with a multi-provider AI assistant that runs a real tool-calling loop - with user confirmation before any record is changed.
Local-LLM Document OCR
Self-hosted OCR that turns scanned PDFs into Markdown using local vision LLMs, so sensitive document content never leaves the machine.
AI Brand and Email Generation
Claude generates and refines a business's brand voice and email templates through a multi-turn conversation, with marker-delimited output contracts for reliable parsing.
AI Spam Detection for Web Forms
Two-stage Claude-powered spam filtering built into the CMS HTTP intercept pipeline, with fail-open safety and full per-request cost logging.
AI-Driven Personal Operating System
25 autonomous agents running around the clock against a Johnny Decimal Obsidian vault - CRM, email, job pipeline, calendar, planner, and content publishing.