I build the systems behind marketing and sales: lead pipelines, email infrastructure, CRM integrations, and the automation that has to keep working when nobody is watching it.
Most of what I build runs in production for a small number of businesses rather than in the open, so a lot of my repositories here are private. Five that are not:
The enrichment and delivery half of my outbound engine, extracted and made source-agnostic: a CSV of leads goes in, verified emails come out, and shipping to the sequencer is report-only unless three separate gates agree. Providers run cheapest first, from SMTP-verified pattern guessing through four free-quota finders to paid lookups behind explicit spend gates with hard monthly caps that fail closed on a corrupt ledger. The do-not-contact list is append-only plain text, checked at ship time rather than run start, because people opt out mid-run.
Python, one dependency. The README's "parts that took the longest to learn" section is the honest changelog of what production taught me.
SEO audits that run on measured evidence rather than on what a model believes about SEO. Thirty-three scripts hit the actual page, headers, sitemap and schema; ten subagents run in parallel over their own domains; and a final verifier deduplicates to root cause, drops anything measurement contradicts, and blocks environment artefacts from being reported as site defects. Forty findings feel thorough and are usually worse than six, because a reader who finds three wrong items stops trusting the rest. GitHub repository SEO is its own lane, since repos rank and READMEs are landing pages.
Python, standard library only. No pip install, no browser, no required API keys.
Skills and a subagent for Claude Code, built around the idea that an agent is most
useful when it is willing to tell you something you did not want to hear. grill-me
interrogates a claim three levels deep instead of stopping at the rehearsed first
answer, and returns a verdict per claim rather than a list of questions.
blast-radius asks what a change touches before anyone notices, which is the
question that actually predicts incidents. integration-forensics is the catalogue
of ways a third-party API lies to you. memory borrows the bi-temporal model from
temporal knowledge graphs so a stored fact can expire instead of quietly going stale.
Markdown, an installer, no dependencies. MIT.
Pokémon TCG analytics. A ridge model fit per rarity cluster estimates each card's expected price from explainable signals, and the gap against the live market flags cards trading above or below fundamentals. Also computes expected value per sealed set. A daily GitHub Actions pipeline pulls from open APIs, accrues its own market history, and redeploys the site.
Python, NumPy (the ridge fit is hand-rolled, not scikit-learn), GitHub Actions, Netlify.
Reply triage for cold-email campaigns. A webhook classifies every inbound reply using OpenAI structured outputs, with deterministic rules running first so the obvious cases never reach the model. Unsubscribes process automatically through the sending platform's API, anything needing a person routes to a channel with full context, and a human approves every outgoing send. Dockerised, CI on every push, 59 tests.
TypeScript, Fastify, OpenAI, Zod, Vitest, Railway.
Outbound infrastructure. Scrapers feeding an eight-provider enrichment and verification waterfall, self-hosted n8n handling delivery across three client workspaces, and every send, reply and bounce logged so the data decides which campaigns keep running. Moving to verified-only sending took bounce rates from double digits down to under one percent.
Web and lead systems for an interstate moving company: a React and TypeScript site with a prerender pipeline that fails the build on SEO regressions, multi-step quote funnels with SMS verification and bot protection, per-channel attribution into the CRM, and abandonment capture that turns a silent drop-off into a follow-up.
The unglamorous half. Append-only ledgers so a retry cannot contact the same person twice, alerts that fire loudly instead of logging quietly, report-only as the default mode for anything that sends, secret scanning on every commit, and deletion tooling that permanently removes a person's data on request to meet CAN-SPAM, GDPR and CCPA requirements.
TypeScript, Python, React, Astro, Node, PostgreSQL and Supabase, n8n, Docker, Railway, GitHub Actions. I develop with Claude Code daily and treat what it writes the way I would treat a junior engineer's code: reviewed, tested, and run in report-only mode before it touches anything real.
B.S. Information Technology, University of Central Florida, December 2026.