Writing · Findings first
What we found, and what it cost.
Write-ups from work that shipped and work that didn't. Each one opens with the finding rather than the setup, and shows the numbers it rests on.
How much rope to give coding agents?
The correct answer balances risk, project complexity, and the gap between your contributors and your coding agents. Three projects, three different answers — including an eight-year-old vision codebase where the rope had to be short.
Read it on Medium (opens medium.com)Everything else
Published on Medium
Delivery
The AI delivery problem. (opens medium.com)
An opinionated take on the lack of correctness in AI delivery, and why the failure is almost never the model.
Reinforcement learning
Using offline reinforcement learning to trial online platform interventions. (opens medium.com)
Developing policies that simulate the impact of incentives on user behaviour, without running the incentive.
Time series classification
Deep learning to predict engagement in online platforms. (opens medium.com)
Can deep learning predict platform churn without extensive domain knowledge? A test of how far the architecture gets you on its own.
Recognise any of it?
Most of what's written up here started as somebody else's stalled project. If a paragraph sounded like yours, that's usually a good reason to send us the questions.