// selected work
I build production systems fast — because I architect the AI and agentic workflows that help build them. The build times below are real, hands-on hours.
// professional
Re-platformed a critical batch-processing workload from in-database Oracle PL/SQL stored procedures to Spring Batch services running on Kubernetes (EKS) — roughly a 4× speedup. Breaking monolithic database jobs into composable, observable, independently deployable steps made the pipeline faster, easier to reason about, and far simpler to operate.
// personal & venture
A complete customer-engagement platform I designed and built solo — multi-tenant by architecture, with marketplace-style lead intake, ticketing, invoicing, and a knowledge base. The interesting part is the AI intake pipeline: it transcribes voicemails into structured leads, triages submitted photos, and drafts review replies — all behind an offline stub so the whole flow demos without live API keys. Built for my own venture over three focused weekends.
The voice front-end to my assistant, Margo: a tiered screener that answers the phone and texts, works out who's calling and why, and passes along only what actually needs me. Real-time telephony wired to an LLM persona, running at the edge.
The marketing site for my venture, KMO Solutions Foundry — React and Vite prerendered to static HTML on Cloudflare Workers. It ships with a build-time brand-voice linter that fails the build on off-brand copy, so the site can never drift from its own voice guide.
Four production-quality demo sites across different verticals — a restaurant, a law firm, a retail shop, and a salon — built from one shared component monorepo to show the CRM in context. Roughly an hour each, because the hard part was built once and reused.
The site you're reading. Hand-built in Astro with a neon-cyberpunk brand and prerendered to Cloudflare Workers — now growing a server-rendered /admin dashboard, gated by Cloudflare Access, to drive Margo. Public and open source.
// ai architecture
My chief-of-staff assistant. One consistent persona that screens my calls, texts, and email, runs my morning brief, and answers questions about my schedule and tasks. Under the hood it's an orchestrator that loads purpose-built subagents just-in-time, a local-first daemon running proactive loops, and an act-low / ask-high approval gate — it handles the safe things itself and drafts-and-holds anything outbound. References are baked in, so it never wastes a turn re-fetching schemas at runtime.
Margo, generalized and given away. Seneschal is the de-personalized framework beneath my own assistant — a local-first, chief-of-staff AI built as a Claude Code skill suite. You bring the persona, the channels, and the data store; it brings the machinery: one orchestrator that picks a mode and runs every turn through an ordered advisor chain, a single act-low / ask-high approval gate, and a resident asyncio daemon that holds a warm chat session and fires reminders on time — subscription-billed, no hosted brain. The system of record is pluggable — Notion, an Obsidian vault, or a plain Markdown folder — behind one seam, and a guided /setup wizard makes it yours one chapter at a time.
How I ship fast without shipping slop. Specifications come first, and a separate verifier certifies every deliverable — the thing that does the work never grades its own homework. The core is a durable task-graph runtime with typed contracts, file-first state, memory, and a self-improvement loop; a local model does the cheap work and escalates to Claude only when it has to. In practice it runs as audit-driven phases: an audit becomes the spec, a briefing hands it off, one pull request per phase, and the verification gate decides what merges.
The tooling that makes multi-agent engineering repeatable: context-loaders that prime a session with the right docs in the right order, a crash-recovery skill that rebuilds progress from git history and a ledger to resume a dead run, an in-flight decision framework, CI bootstrappers, and doc-sync. The tools that build the things that build.