services
The AI feature in your backlog, live in your app in two weeks.
AI integration services for your existing web app: one AI feature in a two-week sprint at $6,000, tested against known answers and shipped behind a flag.
On my own product WhatIsThatBook, the search with a model found the right book in the top five on 68% of test searches. Without the model, the best search reached 29%. Your feature gets the same kind of test before it ships.
Who this is for
SaaS teams with a live web app, paying customers, and an AI feature stuck in the backlog.
Teams with no engineer who has run a model in production: the bill, the wrong answers, the rollback.
TypeScript codebases. I work in TypeScript.
Who it is not for
Business automation in Zapier or n8n.
Advice without code. This is a build sprint.
Founders with no product yet. See MVP development.
Teams that want ChatGPT to call their product. See custom MCP server.
What one sprint can ship
Three example scopes, each based on a feature I run. They are not past client projects.
Search by description. Users type what they remember, and the model finds the record in your data. My book and movie finders work this way.
Questions in plain English. A model writes the SQL, and the answer comes back as a chart, a table, or a map.
A cheaper version of a feature you run. A shorter prompt or a different model, with cost and quality measured before and after.
Proof: AI features I run
These run on my own products, not client work. Each one links to its source.
Tested before it ships. I test changes against 100 known searches stored in the repo. The model pipeline found the right book in the top five on 68% of the 95 answerable cases. The best search without the model reached 29%. Read the test results.
Cheaper to run. One routing change gave about 47% lower model spend in one evening, with almost the same number of results. Median response time went from 7.57 s to 4.18 s. That was the evening of 7 September 2026 against the evening before. Read the numbers.
Off in 14 minutes. One cheaper model failed, and I turned the flag off after 14 minutes. The next one returned books on 89% of searches, against 76% for the old model, over 47 hours. Read the benchmark.
Live and public. WhatMedicaidPays takes a question in plain English and answers with a chart, a table, or a map, over Medicaid payment data from all 50 states.
Words from clients
Justin did a killer job for us, organizing and planning out the first part of our education platform, and honestly, we wouldn't have been able to launch it without him.
He came in and built it from scratch from the very beginning in SvelteKit, he was an integral part of the team as we built up the team around him to be able to support a larger platform like updraft. It was a fantastic experience with Justin.
We would definitely come back to him when we need another extra pair of hands on the SvelteKit side.
Just ten out of ten. Huge kudos.
Thank you, Justin. Bye.
Justin is a delight to work with.
From the start of the project to the end, he made sure to be on par with the scope and even went beyond our expectations.
What started as a business partnership has now evolved into a close relationship between our two businesses, NerdyJoe and Okupter.
How it works
01 / Scope by email. Send me the feature, repo access, and sample inputs with the answers you expect. I reply within one business day. We agree on the scope and on what counts as a right answer.
02 / Start on a sprint Monday. You pay upfront. New sprints start every other Monday.
03 / Mid-sprint sync. One 30-minute call to review the first test results.
04 / Deploy. The feature goes live behind a feature flag. You turn it on for some users or for all of them.
05 / Handover. A short Loom or document, and the test set stays in your repo. You own the code from day one.
Price and refund
$6,000 per two-week sprint. Paid upfront, after we agree on scope and a start date. Model usage is billed by your model provider, on your account. See the full sprint terms.
Something usable after the first sprint. Or I refund it. A first sprint without a usable feature is refunded in full.
FAQ
Here: one AI feature added to a web app that already has users. I build it, test it against known answers, and deploy it behind a flag, in one sprint.
Yes. The feature goes into your code as new routes and components. The rest of the app is not rewritten.
$6,000 per two-week sprint for the build. Model calls are a separate bill from your model provider.
The one that passes your test set at the lowest cost. I test candidates against your known answers before one goes live.
It ships behind a flag. On my own product, I turned a failing model off 14 minutes after I turned it on.