Two ways to work with me, and both start free

I do two kinds of work, and they look unrelated until you see what they share. For founders, I design and build the first version of a product. For AI labs and data partners, I write and review the coding tasks their models are trained and tested on. In both cases the job is the same: decide what matters, build it carefully, and prove it works before anyone pays for the next step.

This post is the long version of what I pinned on X this month. It lays out both offers plainly, with the numbers behind them, so you can decide in a few minutes whether either one is for you.

For founders: a fixed-price MVP, built by one person

Most early products die between the idea and the first usable version. Scope grows, the agency hands you off to a junior, the estimate doubles. I remove those failure points by being the only person on the job and by fixing the price in writing before we start.

I've shipped 22+ products this way, 98% of them on the date we agreed. The case studies on this site show the problem, the decisions and what was delivered for each one. At Ringover I ship AI features inside a call-intelligence platform used by sales teams, so the AI parts of your product are not new ground for me.

For AI labs: coding and evaluation tasks that survive review

Frontier labs and their data partners need realistic coding tasks with tests that cannot be gamed. Most submitted tasks fail review. Terminal-Bench 2.0, for example, accepted 89 of 229 submissions, about 39%, after roughly three reviewer-hours per task. The rejected ones are rarely bad ideas. They are under-specified, or their tests let a shortcut through.

My acceptance rate across this work is 90%, with an average quality rating of 4.89 out of 5. Of the 50 Terminal-Bench-style tasks I wrote for data platforms, 45 were accepted. Each task ships with a pinned Docker environment, a reference solution that is re-run before submission, and hidden tests that fail hard-coding, partial fixes and empty runs.

• Task authoring: Terminal-Bench, SWE-bench and PaperBench-style tasks on real-world problems.
• Task review: a written report on where an agent could shortcut your task and how to close the gap.
• Rubric and grader design: specific, checkable criteria that two reviewers will score the same way.
• Managed delivery: a small, vetted team under one quality gate, when you need volume.

Why one person does both

Writing benchmark tasks taught me to think like a reviewer: assume the solution will be gamed, specify everything the grader checks, and prove the task is solvable before asking anyone to attempt it. Building products taught me the opposite discipline: cut everything that does not serve the one job the user came for. The two meet in the middle. Good tasks and good MVPs are both small, specific and tested.

Everything I post this month follows from that. Monday to Friday you will see short posts on X, Threads and Instagram. Each one has a full write-up here, with sources, so you never have to take a one-line claim at face value.