Case Studies
Selected work, in depth.
Fourteen builds: the Checkmate integrations organization, TargetMobi, and nine independent builds of my own. What the problem was, the decisions that mattered, and what shipped. Client brands are described, not named; happy to walk through specifics on a call.
Platestead: one codebase, a deployment per restaurant brand, two POS providers behind one core
The first-party ordering build, grown into a platform: two demonstration restaurants on one Rails codebase, Square and Clover sandbox integrations behind one core. No real merchant yet.
Read the case study →
WhatsApp Commerce V2: rebuilt for what the demo didn't show
A WhatsApp Catalog + Cart ordering channel, rebuilt after its first version's own logs showed every delivery-status webhook discarded and errors hidden behind HTTP 200s. Webhooks are verified and stored before they are processed, deliveries are processed idempotently, and every reply is tracked to read. Verified with real Meta traffic, the author as the only customer; a production-oriented reference implementation, not a claim of production-scale operation.
Read the case study →
Keepford: an AI phone receptionist where code, not the LLM, decides
The model holds the conversation; application code owns the decisions that matter, from life-safety calls to bookings. Tested with real phone calls by one tester. Built with Python (Keepford); primary stack Ruby on Rails.
Read the case study →
MCP Server: an AI interface over real commerce data, and an independent way to check its answers
An MCP server over 99,441 real, anonymised Olist orders, and a harness that recomputes every answer from the raw data. It found nine correctness defects in tools that already looked right. A live, read-only demo; an engineering evaluation, not a benchmark.
Read the case study →
AI Restaurant Receptionist: the model proposes, the server decides
A Vapi voice ordering agent where Rails owns the order: versioned cart, server-written read-back, and a confirmation gate built from live-call evidence. Deployed, and tested with browser voice calls including one on the production app; no customers.
Read the case study →
Fieldnote: keeping consequential decisions out of the model, then attacking that boundary
A conversational health check-in for community-based participatory research where the model only proposes. An evaluation that never asks the app what is right, and watches what reaches the model, found where the deterministic safety boundary leaked; the fixes went into the mechanisms, not the word list, and later reviews found flaws in the evaluation itself, which were fixed too. The keyword list stops 16 of 60 explicit disclosures. Fictional answers, no participants, no detection rate claimed, and the live model was not evaluated.
Read the case study →
Vrinda House: Stay-Proposal Engine
By design an LLM reads a guest's enquiry and deterministic Ruby decides what a fictional hotel can promise and at what price; the app makes no live model calls. Built by directing Claude-model coding agents. One pre-registered baseline run of a standalone LLM: 3 of 62 broken promises (1 substantive, 2 malformed-only) on synthetic scenarios, scored by the engine's own validator, all on rules or form data the prompt did not give.
Read the case study →
Reviving a 2014 Ruby gem: when not to build an SDK
A Ruby gem I co-wrote in 2014, rebuilt as a focused client for ClickSend's SMS API instead of a second SDK: retries designed to avoid duplicate SMS, ways to reconcile a send whose outcome is unknown, an optional OpenTelemetry companion, and releases published through Trusted Publishing and rebuilt byte for byte. No paid SMS was sent; delivery receipts and a successful cancel were not observed live, so cancelling ships as experimental.
Read the case study →
First-party restaurant ordering, built solo: one POS, two courier networks, five published gems
A complete first-party ordering platform built solo: five open-source gems, a live demo deployment, and the bugs written up in public, unlike the client work below.
Read the case study →
Putting a global burger giant on every leading delivery app
NCR Aloha order injection at enterprise scale, international launches, and a ground-up rebuild that carried the integration across a POS generation change.
Read the case study →
One architecture, five national brands
The enterprise menu-and-launch pipeline behind one of America’s largest restaurant groups — including the largest launch in company history.
Read the case study →
From launch events to a launch pipeline: one brand, four POS platforms
A global fried-chicken brand across Oracle Simphony, NCR Aloha, Xenial Omni, and Panasonic BPOS — brought live continuously, not in one risky event.
Read the case study →
Automated kitchen printing for delivery orders on Clover POS terminals
An on-device Android agent that closed a silent-failure gap on Clover terminals — then was retired cleanly, in one backend change, when Clover shipped its own print API.
Read the case study →
An SMS-based CRM and survey platform for research, NGOs, and healthcare programs
Regional SMS gateway routing, a dedicated healthcare vertical, and support for a 5-year NIH-funded research initiative with UW Madison — owned end-to-end.
Read the case study →