TMT Associate · Programmer
Dual LL.M. Distinction (KCL · Bocconi) · Law by day, code by night
Builder at heart. Was programming in C and Python during law school long before genAI hit, and have never stopped since. Law by day, code by night. During my initial law studies at WU Vienna, I took every AI and legal-tech elective offered and wrote my Master's thesis on AI training data, but what I loved most was happening at night. At École 42, a project-based programming school with no lectures and no professors, I taught myself programming from the ground up. I built a Bash shell replica from scratch (pipes, signals, and all) and two graphical video games, spending many nights debugging programs that crashed unexpectedly and deciphering intricate 1970s C documentation until they finally worked. I kept building throughout law school, including building agentic AI systems at UN, UNIDO, and KCL hackathons. The two threads finally came together when I enrolled in the first-of-its-kind Dual LL.M. in Law, Innovation, and Technology at King's College London and Università Bocconi, where I was supported by a merit-based scholarship and graduated with Distinction. I still code by night, following every release, reading every interesting repo, prototyping constantly with Claude Code and the Agent SDKs. By day, I am a TMT Associate at Wolf Theiss advising international clients on data protection, AI, and cybersecurity. I am also a core member of the firm's legal engineering and digital transformation initiative, where I shape strategy and roadmap development and develop LLM-based solutions (Agent SDKs, workflow orchestration, tools, MCP servers, etc.). I am working closely alongside lawyers, IT, and operations. The hard part is never to prototype, it is building solutions that attorneys actually trust and adopt.
I built an agentic-first legal IDE from scratch: Python/FastAPI backend, SvelteKit frontend, PostgreSQL, and Docker; model-agnostic agentic loop with tool-calling, subagents, multi-step orchestration, and human-in-the-loop review; hybrid BM25 and pgvector RAG retrieval; binary diff document versioning; 6-layer privilege-based permission model.
At the KCL Hackathon, I built an agentic due diligence pipeline built for a Linklaters-sponsored use case. Ingests corporate document sets, extracts metadata (jurisdiction, governing law, parties, dates), and categorises each document into legal domains (corporate, finance, IP/IT, employment, real estate, litigation). Categorised documents are routed to specialised LLM agents that run domain-specific risk assessments, each using role prompting, few-shot examples, chain-of-thought reasoning, and domain-calibrated temperature and top-p settings. Individual agent outputs are verified and merged into a single DD report, surfaced through a dashboard (risks by legal domain, risks by severity) with human oversight built into the final step. Modular solution that is configurable end-to-end via JSON. Python and OpenAI API. Demoed live to a judging panel.
At the UNcode Hackathon, I developed an LLM-based classification system to enhance UNIDO's strategic project programming. The system classifies UNIDO's project portfolio against three frameworks simultaneously: the 17 Sustainable Development Goals, UNIDO's strategic priorities, and its focus areas of expertise. Projects can be selected by ID, country, budget, or duration, with a three-stage data pipeline: structured dataset → PDF extraction → UNIDO API fallback when prior stages return insufficient data. Classification uses a hybrid approach, running generative AI (GPT-3.5-turbo and GPT-4) in parallel with zero-shot classification; because the zero-shot library caps at 10 labels, the 17 SDGs are run in batched elimination rounds and reconciled. Each classification carries an embedded note of justification, directly addressing the traceability gap we identified in pure zero-shot approaches. Outputs consolidated into structured JSON and surfaced through a reactive Dash dashboard with filterable views on country, SDG, priority, focus area, and investment amount. Fully modular via settings.ini and swappable prompts.json/criteria .json files. Python and OpenAI API.
Media Laws
This article analyses the tensions between the GDPR and the AI Act in the context of AI development, focusing on the technical realities of centralised training. It evaluates Federated Learning (FL) as a decentralised, privacy-preserving alternative that advances compliance with data protection principles, unlocks siloed data, and exemplifies “data protection by design”. Yet at the same time, FL complicates compliance with the AI Act’s obligations for high-risk AI systems, in particular data governance, bias mitigation, and robustness, as these were drafted with centralised training in mind. Further analysis details the energy–privacy trade-off inherent in FL. It concludes that while FL provides a credible pathway to trustworthy, human-centric AI development, its distinctive features demand further technical research and either a flexible interpretation of the AI Act’s essential requirements or a dedicated regulatory framework.
Hugging Face Blog Post
Gemini-3 achieved the top spot on most major benchmarks last week, but how well does it know the law? Unfortunately, most model providers don't evaluate on law-specific benchmarks. So while we have a good idea of how good new models are at coding, we are pretty much in the dark about their lawyering abilities. In the HuggingLegal community we were curious how well Gemini 3 performs on some law tasks, so we ran a vibe check.