I came to AI through data. As a Clinical Data Manager I lived in it daily; modeling it was a short leap, and clinical data doesn't tolerate an unverifiable claim — that habit followed me in. When generative AI arrived, I started building LLM tools, first for my own problems, then for other people's.
My rule: problem before solution. I understand it, check whether generative AI genuinely helps, then choose tools. No tool goes in because it's new.
What I'm chasing is the same move each time: find a real problem, build the thing that cracks it. I want to do that with a team, not just for myself.
I build the AI-powered part of a product — the RAG pipeline, the agent workflow, the vision feature — and I build it to survive contact with real users. Every project on this page ships with the same instinct: know when the model is right, and have a real answer for when it's wrong.
Teams shipping an AI feature inside a product that already exists — a Shopify app, a mobile app, an internal tool — who need it production-ready, not a demo. I'm most useful where a wrong answer costs something: regulated or trust-sensitive work, anything customer-facing, anything where "the model said so" isn't good enough on its own.
Small-to-mid SaaS teams adding an AI feature without hiring a whole ML team for it. Founders who've built the product and need the AI part done properly. Anyone in healthcare, clinical, legal, or another regulated space who wants AI but can't afford it hallucinating.
Clinical trial data doesn't tolerate an unverifiable claim — every number needs a source, every change needs a reason. I build AI systems on the same rule: it's why Ayuti cites its sources, why Mersail won't send without approval, why Brezza Viva refuses to generate rather than risk it.
Anyone can summarise a video. Ayuti makes one checkable — every answer it gives is traced back to the transcript line that produced it, across a library that deepens with each channel it watches. The LLM pipeline is multi-provider by design, so one vendor outage doesn't take the product down.
The hard part of AI email isn't the writing — it's that one bad send reaches the whole list at once. So Mersail never sends unsupervised: every brand-matched campaign waits behind an approved preview, and inbound replies are triaged, with safe questions auto-answered and everything else escalated to a human.
Excluding a search term is easy. Excluding one that would have sold is the expensive mistake. So the model never acts alone: every uncertain term gets two independent checks, a term that has ever converted is never blocked, anything in doubt goes to a human review queue, and every block can be undone for 72 hours — with an audit log the database itself refuses to rewrite.
Every other tracker files whatever the coding agent claims. Shipward doesn't: nothing moves until its check actually runs, and git — not the agent, not the board — has final say. A Trust tab even confesses which cards it closed without verifying. It ships as an MCP server; this portfolio runs on it.
Six languages ship from one codebase, Arabic among them — full right-to-left, not just translated strings. The AI is one consented Claude vision call that reads any list, handwritten or printed, and discards the photo immediately after; everything the game runs on — XP, streaks, combos — lives in a pure TypeScript layer with no UI framework tangled in, fully tested.
A generative feature whose best decision was to refuse to generate. Tell it «sto vivendo un lutto» and it assembles a path only from a curated catalogue of passages, always closing in hope, with a guard for acute distress — because a hallucinated scripture reference offered to someone grieving is not a bug you get to ship. The liturgical calendar is computed from first principles rather than fetched.
An AI version of me, running live. Ask it anything a recruiter would ask — background, skills, why I'm moving into AI engineering.
A 30-minute intro call — I'll walk you through my projects and how my clinical data background translates to AI engineering.
