You bring the idea. I build it.It ships.

I was the third developer on the team and stayed a decade — from building the original website and customer dashboards to a company with three offices and a huge call center, serving over a million unique visitors a day.

Third dev on the team

Helped build the original website and customer dashboards, then kept building for ten years as a backend engineer while the company scaled around us.

Scale in the trenches

Over a million unique visitors per day — hands-on with servers running smoking hot, and the database engineering that kept them standing.

First split-test system

Wrote the company's original split-testing platform. It became core to the culture — how decisions got made, company-wide.

APIs for every team

Backend APIs across the business — public pages, customer dashboards, internal tools, and external API integrations.

Marketing engineering

Email marketing infrastructure, attribution tracking, affiliate management — and when the push-notification vendor bills got old, I built the in-house platform so we owned the keys.

Compliance-heavy flows

Email compliance, PII handling, and the guardrails a regulated people-search brand lives or dies by.

When Instant Checkmate took off, we launched a second national brand — and I built the first version of it. Running two brands side by side doubled the surface for experiments: what one property learned, both properties earned.

Built the first version

Wrote the original TruthFinder — a second consumer brand taken from zero to national scale, with everything ICM had taught us baked in from day one.

Maintained at pace

Kept it shipping at an incredible pace and scale — two national properties, one engineering rhythm, no slowdown.

Two-brand experimentation

Parallel brands meant parallel tests — funnels, pricing, and messaging split-tested across both properties at once.

Full-stack build

Database design, backend APIs, and frontend build-out across the subscription business.

Revenue pipelines

Turning paid traffic into subscribers — tracking, attribution, and reporting finance could trust.

A complete event ticketing platform serving the LGBTQ+ community — live at qcal.app and in both app stores. Everything below, from Stripe payouts to App Store review, was designed, built, and shipped by one person.

Payments

Complete Stripe Connect platform — multi-step checkout, orders, refunds, host payouts, platform fees, custom per-host fees and taxes.

Ticketing engine

Per-date tickets for recurring events, tiers and add-ons, batch generation, door check-in scanning, abandoned-cart recovery — and fair-FIFO waitlists that invite in measured waves sized to open seats, so a freed ticket never starts a stampede.

Hosts & teams

Multi-host collaboration with granular permissions, roles and invites, UTM attribution analytics, embeddable venue widgets.

Native mobile

iOS and Android on both stores — push notifications, deep linking, store-compliant account deletion.

AI & extras

Scout, an AI agent that discovers community events from the web and Instagram — plus a built-in page builder powering festival microsites.

The grown-up stuff

A transactional-outbox notification system with consent checked at enqueue AND send, magic-link + OAuth auth, admin moderation, WCAG 2.1 AA audits, a legal risk register, end-to-end purchase-flow tests.

My current love, and the through-line of this whole page: sixteen years after writing Instant Checkmate's first split-test system, A/Bee is that idea grown into a product. It doesn't just test copy — it tests hypotheses about why your audience converts, then keeps what it proves.

A hypothesis engine that compounds

Every test is a claim rooted in proven conversion psychology and tailored to the specific business — each theme targets a different reason that audience might convert. Wins and losses are tracked per site and seed the next round: a winning test teaches you something you can use across the whole funnel.

Runs itself

A cron-driven loop scans every experiment, advances iterations on statistical significance, promotes winners, and immediately generates the next challenger — approval mode or full autopilot.

Guardrails, not vibes

Experiments on proxy metrics get an automatic purchase guardrail with a randomized holdback — the old control stays live on a slice of traffic, and auto-rollback fires if revenue dips. A 'winning' headline can never quietly cost a merchant sales.

An agent pipeline, not a prompt

Challenger variants survive a context digest, a hypothesis critic, a devil's-advocate agent, brand and safety lints, an embedding-based novelty gate, and a judged tournament before they ever see traffic.

Hand-rolled statistics

Two-proportion z-tests with a hand-coded normal CDF, Bayesian Monte Carlo with Beta posteriors, Wilson intervals, and power estimation sized to what a low-traffic store can actually afford to test.

Shopify app with real attribution

Native embedded app: an AI theme reader generates CSS selectors, one script swaps content flicker-free, and the orders webhook ties purchases to variants — dollars per variant, not just clicks.

A hell of a year. Hired to build the public-facing sites, I earned my way into the lab and onto the genetics testing pipeline — software supporting personalized medicine, written by hand, long before AI was around to help. The heaviest testing culture I've ever worked in — the kind of place where you wear a cleanroom suit to walk into the building.

Public sites → the lab

Hired for the web layer, trusted into the clinical one — from the company's public sites to the genetics pipeline itself.

Genetics testing pipeline

Hands-on engineering inside the pipeline powering personalized-medicine testing at one of the country's major clinical labs.

Clinical-grade rigor

Heavy, disciplined code testing — the correctness culture you need when a bug isn't an inconvenience, it's a patient's result.

Miscellany in, delicious out. RecipeMatcher matched the ingredients you already had against every recipe in the index — years before 'search by what's in your fridge' was a standard feature. We grew it into a genuine success; after the founder sold it, it slowly faded away. The engineering was never the problem.

Ingredient-first search

The ePantry: mark what you actually have, match it against every recipe's ingredient list — dinner from what's already in the kitchen.

Hand-rolled everything

No frameworks underneath — search, accounts, favorites, recipe lists, blog, user submissions. Every piece written by hand.

Grown, then sold

Years of iteration grew it into a real success. The founder sold it, and it faded under new ownership — shipped things sometimes outlive their owners.

←1 / 4→

A circus that ended years ago and has not noticed. One page, Spanish first, on a Cloudflare Worker — and because the soundtrack needed making, the project grew an audio editor, which now publishes the finished track and a sheet of timestamps the site reads to throw its own lights in time with the music.

Two languages, one page

Spanish is in the HTML so the page works with JavaScript off; English is a toggle. Written for a border town where half the audience reads each, and translated for Tecate rather than for a dictionary.

A sign that fails properly

Bulbs distributed along an SVG path with a fixed burnt-out pattern and a chase. The lettering, the bulbs and the misregistered red plate all die on one clock, so it reads as a single outage rather than three unrelated glitches — and unlit letters go darker than their surroundings rather than transparent, the way physical letters do.

The flier prints itself

Generated to an unlisted page at exact US Letter size, scaled to fit a phone without reflowing so the crew can check it before a print run. The QR is encoded from scratch — byte mode, Reed–Solomon over GF(256), all eight mask patterns scored — rather than pulled from a library.

Eight fliers, eight URLs

Every poster location gets its own short URL that serves the same page and records the visit in D1. Bots are filtered and the write is deferred, so counting can never delay the page someone came for. A hidden page compares the locations, so the next print run can chase whatever is working.

An audio tool, because the score needed one

Tramoya — Spanish for the rigging and trapdoors behind a stage, the part that makes the ghost appear. Multitrack editing in the browser: clips that split and trim without ever touching the source audio, a separate machine per row, seventeen effects written by hand against the Web Audio API, and MP3 export through LAME in a worker. Projects and audio live in R2, addressed by the hash of their own bytes, so the same track backing two projects is stored once.

The music drives the lights

The editor publishes the track and a cue sheet of timestamps; the site fires its blackouts, red washes and burn-outs off the audio's own clock rather than a page timer, so they land correctly whenever the visitor allows the sound to start. Both Workers share one bucket, so re-cutting the teaser is a publish rather than a deploy.

A personal obsession turned search engine: find the giant ball of twine on your way. The clever part is the economics — content is AI-generated once, then saved forever, so the SEO directory compounds with every search. The 2026 relaunch threw out the old framework and rebuilt it on three Cloudflare primitives — deliberately minimal, and live.

A directory that compounds

Every AI-generated result is saved permanently, keyed by place AND language — the index grows with real use, at the natural pace search engines trust. No million-page dump; a directory that earns its size, one search at a time.

Scored like a critic

Type-aware rubrics, 0–5: landmarks rate on photo-worthiness, fun factor, and historical significance; diners on food quality and wait times — and for anything else, the AI picks the criteria that fit the place.

A concierge on every page

AI visitor tips, heads-up warnings, and an ask-anything chat that keeps conversational state and links you onward to other attractions — chat history lives in localStorage, so it works with no account and no PII.

Rewritten in a night

The whole site rebuilt on three Cloudflare primitives: static assets serve the React app, one Hono Worker serves the API and server-renders every attraction page, and R2 holds the docs. No database, no index, no cron — the key space itself is the directory tree.

Hardened AI surface

User input is armored against prompt injection before it reaches the model, with honeypots and crawl-safety controls around the generation loop.

Built multilingual

Add a language code to an array and the entire UX translates; attraction content follows the visitor's browser language. Live in English, Spanish and French — the rest of the list is one line away.