Writing
On building software, AI systems, and product thinking.
A 200-problem feasibility study showing that small open-weight reasoning models vary enough across instances that a learned adaptive allocator could meaningfully improve the accuracy–compute frontier.