As I’m ramping up my consultancy, I’ve been spending some time learning the tools to build AI apps.
So I built Foodle – an app which dispatches AI agents to find restaurants, bars, and cafes – and plot their recommendations on a map.
Some learnings:
- Documentation is really bad – and even good documentation is out of date <6 months
- AI coding agents are really good at rapid prototyping, and I built a functional prototype <1 day
- Production code still requires a human in the loop. After the functional prototype, I spent ~2 days stabilising and fixing bugs
- AI agents are really expensive to run – and was surprised how often I had to top up my credits while developing
Though my biggest learnings – noticing that:
- Most AI apps are just very thin wrappers around their upstream LLMs – which doesn’t build a very big moat
- And businesses defaulting to AI as opposed to using the best tool for the job. E.g. Foodle would be both much faster (and cheaper) if it were just be a series of API calls
Leave a Reply