About
Hello, and welcome to my portfolio. I'm a Machine Learning Engineer specializing in generative models for visual data (images, video and 3D meshes) and in the infrastructure that trains and serves them.
I'm currently at Unit8, working across AI and data engineering. I built the fine-tuning capability in Darts, the 9.5k-star open-source forecasting library, and gave a large Swiss reinsurer its first consolidated view of how data moves through its Palantir Foundry platform.
The generative work runs deeper. At ORamaVR I built a text- and image-to-3D mesh generation stack end to end: preparing the mesh corpora, co-designing the tokenization, training and evaluating the model, then optimizing inference until a single mesh took under a minute instead of five to ten.
Before that, three years as a research assistant delivering applied ML for industry and public-sector clients, including anomaly detection for the European Space Agency and a multi-agent reinforcement learning system published at IEEE PIMRC.
I hold a B.Sc. in Software Engineering and an M.Sc. in Data Science, with my thesis carried out at Osaka Metropolitan University in Japan. Have a look at the projects below, or use the Resume section for a structured overview of my experience.
Machine Learning Engineer & Software Engineer
- Birthday: 20 March 1996
- City: Biel/Bienne, Switzerland
- Bachelor: Software Engineering
- Email: alain.gysi@outlook.com
- Nationality: Swiss
- Master: Data Science
Skills
I work end to end on machine learning systems: preparing and structuring raw data, designing and training models, and optimizing inference for production. A software engineering background shapes how I build the pipelines and services around them.
Machine Learning
Deep Learning
Generative AI & 3D
Reinforcement Learning
Data & Infrastructure
Languages & Frameworks
Languages
French is my native language, I work daily in English, and I taught myself Japanese to JLPT N2 before carrying out my Master's thesis at Osaka Metropolitan University.
Resume
Education
Master thesis
April 2022 - September 2022
Osaka Metropolitan University, Osaka, Japan
Grade: 6.0 / 6.0
Thesis summary:
- Algorithmic Innovation: Engineered a deep learning framework for conditional, frame-by-frame video generation, synthesizing fluid and temporally coherent motion between disjointed anchor frames.
- Dataset Engineering Architected a specialized data pipeline to curate and preprocess high-fidelity motion data from the Demon Slayer anime, creating a custom dataset that enabled the model to learn and replicate complex, non-linear character movements.
- Research Impact: Developed a solution capable of generating plausible in-between frames that closely align with ground-truth physics and movement.
Master of Science in Engineering (MSE) - Data Science
September 2019 - September 2020
HES-SO Master, Lausanne, Switzerland
Grade: GPA 5.5 / 6.0 | Award: Best GPA
Focus: Advanced statistical modeling, deep learning architectures, and large-scale data processing.
Grade TranscriptBachelor of Science in Computer Science - Software Engineering
September 2016 - September 2019
HE-Arc, Neuchâtel, Switzerland
Grade: GPA 5.6 / 6.0 | Award: Best GPA
Thesis summary: Architected an industrial simulation using Unity ML-Agents and Reinforcement Learning to optimize workshop throughput, enabling agents to autonomously adapt to variable order complexity and reduce production latency.
Grade TranscriptPublications
Unified Zero-Shot Time Series Forecasting: A Darts Foundation
arXiv preprint, 2026
Presents the foundation-model capabilities added to Darts, including the fine-tuning support I built, which lets pretrained forecasters be adapted to a user's own data through layer freezing or end-to-end training.
SOON: Social Network of Machines to Optimize Task Scheduling in Smart Manufacturing
IEEE PIMRC 2021
ieeexplore.ieee.org/document/9569644
Third author, published under my previous name, F. Abdalla. Multi-agent reinforcement learning for production scheduling in a simulated manufacturing workshop, where the trained agents reached the mathematically minimum number of production steps on every hand-computed scenario, including mid-production machine breakdowns and parallel orders of differing length.
Professional Experience
Data & AI Engineer
September 2025 - Present
Unit8, Lausanne / Zürich, Switzerland
- Open Source Contribution: Built the fine-tuning capability in Darts, the 9.5k-star open-source Python forecasting library, so pretrained forecasters can be adapted to new data by freezing layers or fine-tuning end-to-end. Previously they could only be used as-is or trained from scratch.
- Data Lineage & Compliance: Gave a large Swiss reinsurer its first consolidated view of data moving in and out of its self-service Palantir Foundry services, reverse- engineering undocumented Foundry APIs to get at metadata the platform would not expose. Wired it into daily alerts that chase unmapped data sources and revoke access after 30 days.
ML Engineer
September 2023 - August 2025
ORamaVR, Geneva, Switzerland
- Text-to-3D generation: Co-designed an octree-based tokenization system that cut sequence length from 9 tokens per face to 3, one per vertex, tripling the mesh complexity that fits in a fixed context window. Triangle-strip encoding was validated for a further 50-60% on top. Enabled high-precision text-to-3D generation and transformer-based editing.
- Inference Optimization: Rebuilt the test-time pipeline around vLLM with batching and memory optimizations, cutting single-mesh generation from 5-10 minutes to 30-60 seconds and moving generation out of offline batch into interactive range.
- Large-Scale Data Engineering: Designed and implemented a scalable pipeline to process and caption 3D meshes across ShapeNet, MedShapeNet, Objaverse and ShapeTalk, covering scaling, decimation, mesh cleaning and multi-angle rendering, orchestrated with Metaflow and published as versioned HuggingFace datasets.
- Applied Research: Implemented the Cap3D captioning methodology, swapping the paper's BLIP and ChatGPT stages for a locally hosted vision-language model to remove per-caption API cost. Also explored Gaussian splatting as an alternative representation and rejected it, since marching-cubes reconstruction produced meshes too heavy for VR.
- Semantic Scene Synthesis: Enhanced 3D scene generation capabilities by integrating SBERT/CLIP similarity models and OpenAI APIs for intelligent, context-aware mesh retrieval
- Interoperability & Export: Built seamless .usd export pipelines to ensure high-fidelity asset integration between AI-generated outputs and industry-standard engines like Unity and Blender.
- Cloud Training & Experimentation: Ran the full training and evaluation loop on a single 8×A100 node on Azure, using NeMo's multi-GPU training optimizations and Weights & Biases for experiment tracking. Sole engineer on the workstream, presenting results and setting the next round of experiments with a supervising research scientist.
- Advanced ML Implementation: Accelerated R&D cycles by integrating modern frameworks like NeMo and HuggingFace into the core 3D synthesis pipeline, enabling the rapid deployment of novel generative architectures.
Research Assistant
September 2020 - August 2023
Haute Ecole Arc, Neuchâtel, Switzerland
- Full-Lifecycle Development: Led AI projects from initial client requirements and ideation through to deployment of production-ready models.
- Multimodal AI: Developed robust data pipelines for diverse sources, including industrial time-series, computer vision datasets, and natural language.
- Deep Learning Deployment Architected, tuned, and deployed neural networks using PyTorch, TensorFlow, and Ray RLlib, ensuring models met strict performance and latency KPIs.
- Data Engineering: Cleaned and preprocessed data to ensure accuracy, consistency, and develop robust data pipelines to streamline data processing and analysis.
- Collaborative Innovation: Partnered with cross-functional teams to integrate complex ML models into existing enterprise software ecosystems.
Certifications
Hugging Face Deep Reinforcement Learning Course
March 2023
https://huggingface.co/deep-rl-course/
Grade: Excellent
CertificateWinter School in Data Analytics and Machine Learning
Spring 2021
University of Fribourg, Fribourg, Switzerland
Grade: GPA 5.6/6.0
Certificate with gradeProjects
A mix of generative models, reinforcement learning, computer vision and the data infrastructure underneath them, built for clients ranging from the European Space Agency to a viticulture institute. Each page covers what the problem was, what I built, and what came out of it, including the approaches that did not work.
The game jam entries at the end are a different kind of exercise: building something playable from nothing in 48 hours, with friends.
- All
- Deep Learning
- Data Engineering
- Machine Learning
- Reinforcement Learning
- Game Jam
