Hi there, I’m Sergio! 👋😊

AI researcher advancing speech recognition and language understanding at Idiap 🇨🇭🏔️

五つの「気」：勇気・元気・本気・やる気・根気 🌱

Curious about what I do? If we were chatting over coffee ☕ (or mate 🧉) one afternoon and you asked me, I might say:

I'm a Research Associate at Idiap's Speech & Audio Processing group. Never heard of Idiap Research Institute before? (fun fact: PyTorch was born here! 🚀) Located in a lovely valley in the heart of the Swiss Alps 🇨🇭🏔️, Idiap has been a leading institute in AI for decades. My research focuses on the exciting intersection of speech recognition and NLP, stress‑testing and improving LLM‑based ASR architectures (coupling speech encoders with LLMs) and advancing dialog modeling & conversational structure induction for task-oriented dialogues.

For example, in Dialog2Flow we introduced action-driven sentence embeddings that turn multi-turn conversations into explicit workflow graphs. We unified task-oriented dialog datasets and pre-train embeddings that cluster utterances jointly by semantics and communicative intent. As part of this work, I introduced a new supervised contrastive loss for large, fine-grained label spaces — if you’re curious, you can check out its GitHub repo 🛠️.

Before LLMs became mainstream, I worked on interpretable text classification. During my PhD, I developed the SS3 classification model by crafting equations designed to learn important textual features. This model achieved, among all participating teams, the best result in the CLEF eRisk challenge for three consecutive years (2019–2021) and led to the creation of the open-source PySS3 library — if you’re interested, there’s an online demo 🔍 too.

As a researcher, I’ve published and reviewed for top-tier NLP and speech venues, including ACL, EMNLP, INTERSPEECH, and ICASSP. You can check out the full list of my publications on Google Scholar.

Recent highlights ✨

Johns Hopkins University's JSALT 2023 & 2025: contributed to the Automatic design of conversational models and gave lectures on, and contributed to, synthetic dialog generation with LLMs as a senior researcher member — you can also see me presenting on the final day 🙈 on YouTube 🎥.
Winner 😎 of the ICC 2024 AI Hackathon: call-center workflow assistant 🧑‍💻.

About me and my background? I’m an Argentinian 🇦🇷🧉 who's been coding since childhood (20 plus years of experience now! 👶💻), and curiosity always kept pulling me deeper and deeper. Here are a few projects I invested a lot of work in the past and I’m especially proud of:

T‑World: an open-source 3D platform for AI research and education where you can code and test your intelligent agents. The platform and the 3D engine are coded completely in pure JavaScript so it runs natively inside the browser without any plugins or server. You can code agents in any language, and I’ve included examples like Q-Learning-based reinforcement learning agents and A*-based problem solvers. You can even play against them in your browser 🎮, Check it out! — click on "Run T-World!" at the bottom and run the first environment of the list 😎
As a Computer Science student at UNSL, I also created a C-inspired programming language, from scratch, for people learning to code. I designed its context-free grammar, wrote a recursive-descent parser in C++, and implemented a stack-based virtual processor as its interpreter and the compiler logic for it. Later, I explored compiling directly to machine code and built a library to generate valid Windows .exe (PE) files from raw x86 machine code — I paused its integration to the compiler after switching my professional life fully towards AI research 😅
For fun, curiosity and learning, I’ve loved to implement many things from scratch. For instance, communication protocols (HTTP, TCP, UDP, WebSocket, etc.) based on their RFCs. For example, I coded a custom WebSocket proxy server in C to connect external agent programs to the in-browser 3D simulation in T‑World. I also implemented small physics engines, the math behind 3D and 2D rendering, and simple-reflex intelligent agents with interesting emergent behaviors 🙈

On the side, in the past I’ve ventured into 2D and 3D animation (some pieces even aired on MTV Latinoamérica 📺), ethical hacking, and teaching. I also enjoy playing classical guitar 🎵 and learning Japanese 🇯🇵 — at one point, I reached the milestone of reading around 1000 kanji!

If you’d like to connect, feel free to reach out at sergio.burdisso@gmail.com ✉️. I’m friendly, passionate, and open-minded, so don't hesitate 💪. Alternatively, you can connect with me on LinkedIn.