Senior Product Manager · AI-driven products · Bratislava

I don't trust an AI feature until I can tell you the number that would make me kill it.

4+ years in product across fintech, insurance and banking (Vacuumlabs) - now focused on how AI actually gets designed, tested and shipped inside real products. This site collects the AI work I can show in full: what I built, what I owned, and what I'd still call unproven.

Kristína Magyarová · Senior Product Manager · linkedin.com/in/kristymagyarova
6AI initiatives
led or shaped
4+years
in product
10+products & platforms
shipped
6industries

case studies

The two pieces of AI work I can show completely - built end to end, with the reasoning, the mistakes, and the numbers included.

Eval case study · built with n8n

The eval said 0.96. The product was dropping every number in the story.

I built a golden-set, LLM-as-judge evaluation layer for a personal news-digest pipeline - then validated the judge blind against my own labels. It caught a truncation bug that had been quietly cutting real accuracy from 0.96 to roughly 0.17, and measured a ±0.04 noise floor that most "prompt improvements" never clear.

Read the case study →

iOS app · built with Claude Code

56 commits, zero Swift experience - the gym app I actually use

I'm not a programmer. I wrote the spec, made the product and data-model decisions, and reviewed every change - Claude Code wrote and committed the Swift. This is what that collaboration looked like in practice, including the design pass we redid three times and the feature I deliberately didn't build.

Read the case study →

more AI work

Work I've delivered through my day-to-day role - summarized here where client confidentiality applies, ordered by how hands-on I was.

AI-Driven Risk Commentary Engine - InsureTech broker, CEE Designed and iteratively tested the prompts powering an AI-driven risk-commentary engine that analyzes historical claims trends and generates recommendations; handed to engineering for implementation.
AI Chatbot Initiative - internal, my own initiative I spotted the opportunity, designed it, and built it in n8n myself, for our marketing team. Currently in active development - I own the roadmap and work directly with engineering on integration.
AI Insurance Proposal Comparison - InsureTech broker, CEE, PoC Designed the UX and Figma prototype for an AI assistant that extracts and compares insurer proposals into a clear comparison table with a recommended offer. The underlying multi-agent extraction and comparison engine was built by the engineering team.
Legal AI Tool & Chatbot - Legaltech company Contributed to a legal AI tool and virtual assistant built on LLMs - planning, key features, and scoring answer quality against the client's historical responses.