AI is useful when it does one real job well. It is expensive noise when it is bolted on for the sake of it.
I build AI features and agents around a specific task: answering customer questions from your own documents, understanding a photo of a meal, enriching leads, or moving data between tools without someone copying and pasting all day.
Things I have built
- A RAG chatbot SaaS that answers from uploaded knowledge, with subscription billing. See the project.
- AI food recognition in a mobile nutrition app that estimates what is on the plate. See Superheart.
- AI meal logging from photos, voice and restaurant menus in my own app, MacroScout.
- LinkedIn automation extensions that extract, enrich and organise leads. See JarvisReach.
How I keep agents reliable
Grounding. Answers come from your data, with sources, instead of the model’s memory.
Limits. The agent can only use the tools and actions you approve.
Visibility. Logs of what it did and why, so a human can review and correct it.
Cost control. The smallest model that does the job, caching where possible, and usage caps.
