Writing
On building production AI — vision, LLMs, agents, and infrastructure.
A UK client selling social media growth services needed a real store, not a landing page. Customers had to browse services across Instagram, TikTok, Facebook and YouTube, pay securely, and actually get their orders fulfilled. And the client needed to run the whole thing day to day without a developer babysitting it.
Language models sound just as confident when they're wrong about numbers as when they're right. If you're going to trust one with financial reasoning, an overall accuracy score isn't enough — you need to know where it slips and why.
Speech-to-text, an LLM, retrieval, and text-to-speech in one real-time loop — so it talks back with answers grounded in your content, not made up.
Models sound just as sure when they're wrong about numbers as when they're right. FinEval is 1,000+ questions and an automated judge to see where they slip.
An agentic AI setup — RAG plus tool-using agents — took a manual finance workflow from two weeks down to about a day, without giving up accuracy.
Manual annotation doesn't scale. A vision-model pipeline let us skip hiring five annotators and saved about $24K a year.
Always-on GPUs are the fastest way to burn a cloud budget. Here's the split that cut ours by more than half without hurting latency.