I embed with the operator and ship the system into production.

ForeRun

Subscription iOS app. Designed, built, and launched solo.

A parent speaks a brain dump or photographs a school flyer. An LLM pipeline turns that unstructured input into dated tasks, assigned to the right child, on a shared household board.

• Claude Haiku 4.5 by default, escalating to Sonnet 4.6 only when a photo comes back empty, so the expensive model is paid for only when the cheap one fails.
• Speech recognition and image downscaling run on device: voice never leaves the phone, and uploads stay under the payload cap.
• One forced tool call, strict server side validation, and a parent confirmation step before anything lands on the board.
• A 15 case eval suite, including a prompt injection case, fails below 0.8 recall. No prompt or model change ships without passing it.

BoothOps

iPad photo booth platform. A study in parallel agent delivery.

Four Claude agent lanes ran at once over two days: an iOS camera feature, a live slideshow, and web and iPad audit fixes. Each lane had its own git worktree and its own private Supabase stack, and lanes merged each other's finished work before landing.

• Reviewer fix rounds on every lane. Review caught a merge that had silently disabled a test.
• Generated database types regenerated from each private stack, never merged by hand.
• Mutation testing (29 of 29 mutants killed) and flaky tests stabilized to 10 of 10 runs.
• Final gate: typecheck, lint, build, browser and native suites, with native tests run under Address and Thread Sanitizers on a fresh stack.

Industrial quote intake

Built for a transformer supplier after discovery with their CTO.

Their sales team quotes from messy customer emails and PDF spec sheets. The demo turns either one into structured quote fields that map to their existing quoting layer over NetSuite: kVA, voltages, phase, unit type, quantity, and delivery.

• Uncertain fields are flagged for a technician rather than guessed. Editing a field reruns rules and matching instantly, never the model.
• Matches each request against in stock units and drafts the customer reply.
• Built to their stated goal of systems that never need managing: one source of truth, fails safe, no new logins.

Georgia Diesel Works

Catalog and shipping data rebuild for a dropship parts retailer.

A scraped catalog of 10,083 products and 16,952 variants was losing money on every heavy order. One 76 lb exhaust cost $75 to ship against $51.95 collected. I rebuilt the data, then rebuilt shipping on top of it.

• Imputed missing weights from the catalog's own data with a deliberate round up bias, so any error overcharges instead of losing money.
• Validated before trusting it: 434 of 440 weights matched a vendor dealer feed exactly, and a hold out test on 2,917 items undercharged only 2.2% of the time.
• Measured dimensional weight instead of padding everything. Only radiators were a real problem (2.46x), so they got their own tier.
• Moved fixes to the Admin API after Shopify's bulk tools silently failed and a stale CSV import silently reverted earlier work.
• Archived 984 noncompliant products while deliberately keeping about $268K of legal repair parts, without deleting a single record.

TrueARV

Property valuation engine for Upstate SC investors.

Reconciles two property data providers into one comp adjusted value with a confidence band. Before any user saw a number, I backtested both providers against real closings, and the measured error reshaped the confidence model.

Dispatch Layer

Route optimization for rental delivery fleets.

Models a live logistics operation as a constraint problem: delivery and collection pairs, setup time windows, per truck capacity, and fleet load balancing. House level geocoding handles new construction addresses other providers miss. It went back to paper twice before the routing respected setup windows the way operators actually think.

Roof Rangers lead routing

Live lead pipeline for a roofing client, plus an AI layer built and tested for it.

Facebook and web leads trigger an instant SMS alert to the owner. Web leads without a consent certificate are held for TCPA compliance. When carrier filtering blocked early alerts, I traced it to number registration and fixed it.

The next layer, a Claude driven engine that qualifies, books, and follows up with each lead (Sonnet 4.6 with a Haiku fallback, quiet hours, opt out, human handoff), is built and tested end to end, not yet live.

Kalshi weather trading system

Autonomous prediction market system, built from scratch.

Ingests ECMWF weather ensemble forecasts across eight city markets, converts them to calibrated probabilities, sizes positions by Kelly criterion, and reports through Telegram. Live performance is instrumented, and measured edge decides which markets it is allowed to trade.