Gigax

<p></p><video-component src="https://www.youtube.com/embed/B-Hw2tmYBwc"></video-component><p></p><p><strong>Description</strong><br>I designed and developed the full backend of <em>Entropedia</em>, a real-time card game playable in the browser.<br></p><p><strong>Features:</strong></p><ul><li><p>Authentication and session management via WebSockets.</p></li><li><p>A real-time <strong>PvP matchmaking system</strong> (queue management, connection tracking, live stats).</p></li><li><p>Orchestration of battle phases (draws, turns, rounds, scoring, KO logic).</p></li><li><p>Integration of an <strong>ML model with constrained outputs</strong>, ensuring every generated card respected a strict schema (avoiding hallucinations).</p></li><li><p>Real-time API communication with the <strong>Godot frontend</strong>, ensuring perfect synchronization between players.</p></li></ul><p><br><strong>Impact</strong><br>The project lasted <strong>8 months</strong> in agile mode, with 60 playtesters and many hours of gameplay ! Today, the game is fully playable in the browser (demo available at <a href="http://entropedia.xyz" rel="noopener noreferrer nofollow" target="_blank">entropedia.xyz</a>).</p><p></p>

Developed the full backend for a multiplayer card game where an ML model creates new cards on the fly. Features WebSocket matchmaking, and schema-constrained card generation. Playable in the browser at entropedia.xyz.

LLMs / Large language models

Backend Development

Python

Real-Time Card Game with ML-Generated Cards

Philharmonie de Paris

<p></p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmfgkk0imfshe5rdiaz180t38.jpg"><div class="image-overlay"></div></div></div></div><p><strong>Client’s Issue:</strong><br>The Philharmonie de Paris wanted to make over 1000 hours of recorded conference audio searchable and accessible. The recordings featured multiple speakers, varying audio quality, and specialized vocabulary.</p><div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmfgkk0iofsyp5xf97sjjgxvu.jpg"><div class="image-overlay"></div></div></div></div><p><strong>Solution:</strong><br>I built and deployed a Python-based audio transcription pipeline:</p><ul><li><p>Raw recordings were processed using OpenAI’s Whisper model for STT,</p></li><li><p>The transcription were then processed by language models to perform typo correction and entity extraction (composers, titles, genres...)</p></li><li><p>The cleaned transcripts were embedded and integrated into a search engine for fast, semantic retrieval.<br></p></li></ul><p><strong>Impact:</strong><br>The Philharmonie has released a closed beta version of the project, allowing full-text search across its conference archive. This was showcased to other musical institutions in Europe (Théâtre National de Strasbourg) that are now reproducing this stack on their own databases.</p>

Designed a pipeline to transcribe, clean, and structure conference recordings using Whisper and LLMs. The processed text was embedded into a semantic search engine, now used in beta by the Philharmonie.

Machine Learning

"The pipeline delivered exactly what we needed — accurate, self-hosted transcription and search that brought our entire conference archive to life."

Making 1000 Hours of Concert Talks Searchable Using Embeddings and Language Models

HuggingFace

<div class="image-container"><div class="image-row-wrapper"><div class="image-wrapper"><img src="https://cdn.collective.work/cmfgk2vgwg3wx5udud07s391y.jpeg"><div class="image-overlay"></div></div></div></div><p><strong>Client’s Issue:</strong><br><a target="_blank" rel="noopener noreferrer nofollow" href="https://huggingface.co/">Hugging Face</a> is a leading cloud company specialized in Machine Learning. They wanted to promote their cloud platform, called <em>Spaces</em>. <br>They needed a compelling video game demo running in the browser to attract game developers.<br></p><p><strong>Solution:</strong><br>Gigax, my company, had just released an <a target="_blank" rel="noopener noreferrer nofollow" href="https://github.com/gigaxgames/gigax">open-source GitHub repo</a> helping game devs integrate clever NPCs in their games using Language Models.</p><p></p><p>HuggingFace commissioned us to plug our APIs into a live, browser-based NPC demo on Spaces. We collaborated with their DevRel team and a game engine company <a target="_blank" rel="noopener noreferrer nofollow" href="https://cu.bzh/">Cubzh</a> to create a live 3D environment running in the browser, where users could talk to various NPCs and then customize them with minimal code.</p><p></p><p>To achieve this, my team mastered several of the language models' "tools of the trade":</p><ul><li><p>Fine-tuning open source models with LoRA, using our own training datasets</p></li><li><p>Self-hosting our models on cloud GPUs with parallelization using <a target="_blank" rel="noopener noreferrer nofollow" href="https://github.com/vllm-project/vllm/">vLLM</a></p></li><li><p>Structured outputs generation with <a target="_blank" rel="noopener noreferrer nofollow" href="https://github.com/dottxt-ai/outlines">Outlines</a> to force the models to follow our custom format</p></li><li><p>Monitoring model performance with "LM-as-a-judge" techniques using <a target="_blank" rel="noopener noreferrer nofollow" href="https://giskard.ai/">Giskard</a> and <a target="_blank" rel="noopener noreferrer nofollow" href="https://langfuse.com/">LangFuse</a><br></p></li></ul><p><strong>Impact:</strong><br>- <a target="_blank" rel="noopener noreferrer nofollow" href="https://huggingface.co/blog/npc-gigax-cubzh">Official blog post</a> relayed on HuggingFace's social media accounts (2M+ reach) <br>- 2K+ subscribers to their <a target="_blank" rel="noopener noreferrer nofollow" href="https://huggingface.co/learn/ml-games-course/unit3/introduction">video game tutorial series</a></p>

Built a live browser demo where players could talk to and customize game characters (NPCs) powered by language models. The project was commissioned by HuggingFace and showcased on their platform Spaces.

A technically excellent, creative integration that resonated with our community and showcased Spaces at its best.

Interactive NPCs in the Browser using language models

ML consultant & Engineer

Trained open-source models and built ML agents for the gaming industry:
• 300+ stars on GitHub => https://github.com/GigaxGames/gigax 
• Demo commissioned by Hugging Face, featured in their blog and courses => https://huggingface.co/blog/npc-gigax-cubzh

Founder

Improving musical databases using AI tools: speech to text, semantic search, and more :)

NLP Consultant

Developed tools and scrapers to update and monitor a database of several million images.

Backend developer

Intern then first hire in a lead generation startup acquired a year later.
Managed monthly delivery for 10+ clients including Google and Waze.

Tristan Deborde

ML consultant & Engineer

About me

Work experience

Founder

NLP Consultant

Backend developer

Project Manager

Education

42 Paris

emlyon business school

Portfolio

Real-Time Card Game with ML-Generated Cards

Gigax

Making 1000 Hours of Concert Talks Searchable Using Embeddings and Language Models

Philharmonie de Paris

Interactive NPCs in the Browser using language models

HuggingFace

Talents

Recruiters

About