ABOUT ME
I am an AI engineer based in Vienna, Austria, with a background in mathematics, working on machine learning since 2016 in academia and industry. These days my focus is on AI agents: building the platforms they run on, and making them reliable enough to automate real work.
At Unitary I went from building the data and evaluation foundations behind content-moderation models to co-founding the platform every customer AI agent is built and run on, and driving company-wide adoption of AI coding agents. Before that I built ML systems for a leading cybersecurity company, and did research at the Technical University of Vienna and Priberam.
I like to work on projects where a little bit of everything is needed. In my spare time I build tooling for coding agents, like agent-pawl.
Outside of work I’m passionate about music, and in particular metal (check out some of my work), and also video-games.
- AI Agents
- Large Language Models
- AI for Software Engineering
- Natural Language Processing
- Cyber Security
- Music
MSc in Applied Mathematics, 2015
Técnico Lisboa
BSc in Applied Mathematics and Computation, 2013
Técnico Lisboa
EXPERIENCE
Technical lead for how Unitary builds AI agents: the platform they run on, the workflow format they’re generated from, and the decision engine they use.
- Co-founded Agent Factory, the platform every Unitary customer agent is built and run on, drastically cutting new-customer delivery effort.
- Spearheaded the workflow design format, a structured spec of each customer’s process that serves as the single source of truth agents are generated from.
- Led the team that built a shared decision engine with automated prompt optimisation, now standard for every new customer with automated decisions.
- Drove company-wide adoption of AI coding agents, from a few early adopters to every engineer.
Led new-customer AI deliveries and helped shape Unitary’s move from ML classifiers to AI agents.
- Authored Automation Rate as the company’s core measure of value, and the product vision behind its pivot to AI agents.
- Led the team that built Trustpilot’s AI agents, cutting manual moderation effort by 78% and making decisions 4x faster (case study).
- Took a global gaming company’s content moderation from manual review to its first production LLM system.
- Delivered AI automation for four more customers, including +19% customer satisfaction for Attune.
Built the data and evaluation foundations behind Unitary’s content-moderation models.
- Built the ML data pipeline platform (Spark, Delta Lake, AWS) and one-command model training.
- Set the evaluation standards for custom customer models: gold-standard test sets and traffic-realistic benchmarks that model releases were measured against.
- Led the internal team that delivered a resale marketplace’s image-moderation classifier.
Built the machine-learning systems that classified the web for CYAN’s content-filtering products.
- Led modelling for a multimodal (text and image) website classifier covering 100+ languages, which found over 10x more harmful sites than the system it replaced.
- Built the Spark/Databricks pipelines on AWS feeding it ~100k websites a day, plus a serverless scraper fetching millions of pages daily.
- Built a domain-embedding model over DNS traffic for phishing detection.
- Stepped in as interim ML team lead (4 engineers), and represented CYAN in an FFG-funded research collaboration with SBA Research and Politecnico di Torino.
I was part of the BigDAMA project, a fundamental research project focused on big data analytics for network traffic monitoring and analysis.
My research topics are mostly related to how to represent network traffic for detecting attacks at the network level (i.e., which features to use). Experimented with both classical features (what have people been using and why?), and feature learning approaches using Deep Learning techniques, and in particular representing traffic in 2-dimensional spaces.
I was part of the SUMMA project, an H2020 project for media monitoring with multiple industry and academic international partners.
I specifically tackled the problem of named entity recognition in which, given a text with some entity mentions (e.g., David and Victoria), the goal is to find the corresponding entities in Wikipedia (e.g., David Beckham and Victoria Beckham). We used both classical methods (such as SVM) and modern Deep Learning approaches.
