Everything I've built, shipped, broken and competed in — ordered from the work I'm proudest of down to the early experiments that got me here. Most of it is self-taught through projects across very different domains.
My first professional role. Tasked with designing and prototyping an AI-powered voice-caller assistant for customer service — owning it from research all the way to a working, deployed prototype.
An actively run archive maintained and updated by my personal AI agents. It tracks Hugging Face models, datasets, papers, and GitHub AI projects, while also acting as a resource my assistant can query to understand AI as a topic and answer with stronger context.
Clavis is my private AI assistant layer: a serious, tool-using agent that helps maintain this website, operate the AI Research Archive, work with code and Git, and turn ongoing research into usable context.
A 48-hour build sprint focused on improving how software is built. The challenge: get hands-on with IBM Bob — an AI that understands your codebase and works with real context — and ship a complete product from scratch. The hackathon ran in May 2026 with a $10,000 prize pool.
Prism reimagines how users interact with web pages by letting AI apply seamless stylistic and structural changes directly in the browser. With predefined modes and prompt-based requests, it can switch elements, patterns, colors, backgrounds, borders, and broader visual structure to match each user's style while keeping the original content intact. Compression techniques help redesign even complex pages efficiently. IBM Bob was a core part of the workflow, helping shape a large portion of the backend architecture and code during the 48-hour build.
My submission to OpenAI's parameter-golf challenge: a novel Encode–Think–Decode (ETD) transformer using U-Net skip connections and Int5 GPTQ quantization, reaching 1.1169 BPB under a strict 16MB artifact limit and 10-minute training budget on 8×H100 GPUs.
A web tool I built to parse and analyze training logs from the Parameter Golf runs. It plots metric trajectories over the strict 10-minute budget, making BPB drops and parameter efficiency easy to track. Built to accelerate my own research and open for anyone to use.
An AI & Business Data hackathon tackling real-world challenges with ML and large-scale datasets. Featured an ML tournament for industry classification, partner side-quests (Banca Transilvania), and access to Veridion's 115M+ company database. Grand prize: €5,000.
Went beyond using AI APIs to actively developing and training models — fine-tuned open-source LLMs for specific tasks, trained image-generation models for custom outputs, and built RAG pipelines connecting LLMs to structured data.
Penetration testing and bug bounty programs. Core approach: business logic exploitation — probing endpoints, reverse-engineering intended server-side behavior through varied payloads, and using legitimate functionality in unintended ways. Also OWASP methodology and network recon. Specifics under NDA.
Built automation scripts, web scrapers, and backend services in Python. Started managing my own Linux servers — Nginx config, reverse proxying, load balancing across upstreams, SSL with Let's Encrypt, and Bash deployment scripts.
Learned HTML, CSS, and JavaScript, then moved into responsive layouts with Tailwind CSS and DOM manipulation.
First experience shipping a complete product. Learned C# scripting, component-based architecture, physics systems, and UI design in Unity — and started using Git for version control.
Started by writing game logic in raw C/C++ — console programs, file I/O, basic data structures. Built a foundation in low-level thinking, memory management, and pointers before ever touching an engine.