Working builds
Three kinds of proof live here, from deepest to lightest: full feature assessments produced by my Opportunity Engine, quick prototypes built for specific roles, and quick experiments that test one sharp treatment. Each takes a real problem and ships something honest, so the thinking is visible instead of described.
🔒 Private and unlisted. These pages are shareable by link but are kept off my site's navigation and out of search. Concept work, not affiliated with the companies named.
Feature assessments
Full assessments of a feature idea, run through a true discovery and definition process across multiple disciplines before any build is committed to. The most realistic of the three: this is what it looks like when an organization moves the PM function up the stack. Produced by the Opportunity Engine and published with the unflattering findings intact.
Customizable navigation drawer
I gave a 15-agent pipeline a two-paragraph idea and two screenshots of the Claude app's menu. It cut a third of my request, built a working prototype, filed 21 risks, and told me not to fund the build yet. Step through the run stage by stage, or skip straight to the finished assessment.
Quick prototypes
Built from user research, a defined problem, and one curated solution that still aligns to the company's overall goals, without yet weighing every function's risks and concerns. The right first step for earning buy-in before going deeper on the wider impacts.
Fullscript Practitioner Activation
Two practitioners sign up on the same day. One has a patient waiting, the other is deciding whether Fullscript fits their practice. The sign-up asks neither of them why they came, so both get the same screens. This asks once and sends them down different paths: one racing to a first plan while verification runs underneath, the other building on a practice patient with no credentials at all. Includes what was verified, what could not be, and the three risks worth testing first.
Marketing Agent
Jobber's marketing tools are a pile of separate signals: reviews, job history, local conditions, past results. This is the agent that combines them, reading every signal to identify the few campaigns actually worth running, with the cost-benefit worked out, and getting sharper each time. The pro keeps doing the work and keeps the send button. The agent does the marketing thinking.
Move Readiness
The listing-page calculators tell you what a home costs to own. They assume the down payment is already cash. This models the whole move, both transactions, so a homeowner sees what it would really take, from the deposit due on offer night to the monthly payment, and where every dollar comes from.
Regulars
Owner's marketing runs on the calendar: Mother's Day, the Super Bowl, an abandoned cart. This organizes it around the guest instead. Five stages from first-timer to regular, learned from each restaurant's own order rhythm, one explicit campaign per stage (and two stages deliberately left alone), points instead of discounts as the default offer, and every result measured against a holdout.
Margin Manager
A forward-looking co-owner for food cost and pricing inside Owner's operator app: sees ingredient price spikes coming, recommends margin-safe price moves weighed against how customers will react, and gets new items sellable with an instant photo.
Upsize This Meal
Turns a common churn reason ("portions are too small for us") into a reason to stay: surfaces which weekly meals can be bulked up at home with a few staples, while holding the recipe's flavour.
Quick experiments
Smaller items that are riskier, or that surface one unique treatment. Built for quick alignment on ambiguous ideas that need explicit testing, loosely grounded in user feedback, and cheap enough to throw away if the test says no.
Made For How You Wear It
A one-module product-page experiment, built in a single working session. Mejuri's biggest price decision, the same design in silver, vermeil or solid gold, is made through two unlabeled swatches. This asks how the piece will actually be worn and recommends the material honestly, both up and down the price ladder. Includes the experiment design: primary metric, guardrails, rollout.
Build logs
Not concepts. Things I actually built and still run, documented as-built — including the parts that went wrong and what the documentation got wrong.
Self-hosting an AI agent on a dead MacBook
A 2015 MacBook Pro that Apple stopped patching in 2024, wiped and rebuilt as a headless agent I talk to over Telegram. The build is less about installation than about deciding what an autonomous agent is not allowed to do: no shell, no browser, filesystem confined, one person on the allowlist. Includes the seven places the documentation was wrong.