I build AI harnesses that are predictable and fail safely.
Not every problem is LLM-shaped
Attention Is All You Need and early ChatGPT were huge breakthroughs, and language models are impressively flexible. However, some problems are better handled with coded logic, such as precise data manipulation.
It's important to be thoughtful: make the right set of model calls, provide targeted context, and build a limited set of effective tools.
Transparency
I believe in being responsible and open about AI use. Humans should know when and how LLMs were used - to write something, to make a decision. It's a matter of respect; I'm all for offloading appropriate tasks, but I want to know how much human effort was put in, so I can put in a similar amount in return.
Guardrails & fallbacks
LLMs are stochastic models, so it's important to plan for failure: validate output, reduce the blast radius of mistakes, and make sure that when things go wrong, the system (or a human!) fixes them.
Coding
LLMs are great for prototyping, translating, researching, and finding issues. That said, I find that building maintainable, extendable, and reliable systems still requires close oversight.