A 1967 Math Paper Just Solved AI's $100 Million Problem
DeepSeek's mHC reduces signal amplification from 3000x to 1.6x using Sinkhorn-Knopp, a 59-year-old algorithm. Training stability solved.
Read on MediumDelanoe Pirard
Reinforcement Learning · Computer Vision · Robotics · Brussels
Building intelligent automation systems and bridging the gap between research and real-world applications.
About Me
Professional Background
I'm an AI engineer and researcher with more than 9 years of experience blending software engineering, data science, and deep learning. Right now, I'm specializing as an AI Automation Engineer, working on smart systems that automate complex tasks, while remaining an active AI researcher exploring new frontiers in the field.
Technical Expertise & Interests
I love building robust, reliable AI models that actually work in real life. Recently, I've been focusing on robotics projects using reinforcement learning, realistic simulations with Mujoco and IsaacSim, and computer vision techniques to make robots smarter and more precise.
Academic Journey
I earned my Master's in Bioinformatics and Modeling from Université Libre de Bruxelles. My master thesis focused on predicting protein melting temperatures using advanced NLP methods, specifically fine-tuning ProteinBERT. I was also involved in developing computer vision models to detect pancreatic cancer from histological images.
Beyond Engineering & Research
When I'm not at work, you might find me replicating AI research papers to stay sharp, composing electronic music (something I've been passionate about for over a decade) or diving into philosophical readings, another deep interest of mine.
0
Years in Tech & AI
0
Years as AI Automation Engineer
0
AI & ML Projects
0
Professional Certifications
Skills & Expertise
Skills overview: Machine Learning 80%, Deep Learning 75%, Reinforcement Learning 55%, Computer Vision 55%, NLP 50%, Data Engineering 70%
Frameworks & Tools
Research Skills
Data Analysis
Statistical analysis, hypothesis testing, and experimental design
Literature Review
Comprehensive analysis of research papers and state-of-the-art methods
Technical Writing
Redaction of complete technical documentations and research papers
Collaboration
Cross-functional teamwork with researchers from diverse backgrounds
Research Areas
Deep Learning & Automation
I develop practical deep learning models and AI agents for intelligent automation, enhancing real-time decision-making in robotics and industry.
Bioinformatics
I apply advanced NLP and machine learning techniques to tackle real-world bioinformatics challenges, such as predicting protein stability and analyzing biological sequences.
Computer Vision
My work involves developing accurate and reliable computer vision algorithms, particularly for biomedical applications such as early and precise cancer detection.
Reinforcement Learning & Simulation
I design advanced reinforcement learning methods combined with realistic simulations (like Mujoco and IsaacSim) to enhance robotic performance and autonomy.
Featured Projects
📄 Paper Replication
Titans PyTorch MLX
Complete PyTorch and MLX implementation of Google's Titans architecture with Neural Long-term Memory. Supports MAC, MAG, MAL variants with Flash Attention 2 and Metal kernels.
View Code
MLX-MASt3R
Native MLX implementation of MASt3R for 3D reconstruction on Apple Silicon. 1.87x faster than PyTorch MPS with custom Metal kernels for fused RoPE 2D, bilinear upsampling, and grid sampling.
View Code
Awesome Depth Anything 3
Production-optimized fork of ByteDance's Depth Anything 3 with 200x faster cached model loading, adaptive batching, Apple Silicon optimization, and PyPI distribution.
View Code
Lang-Efficient-SAM
Text-prompted object segmentation model combining EfficientSAM and GroundingDINO for efficient, language-guided image segmentation with minimal computational overhead.
View Code📄 Paper Replication
ProteinBERT PyTorch Implementation
Full PyTorch replication of ProteinBERT (Brandes et al., 2022) for protein sequence analysis using self-supervised transformers and transfer learning on biological sequences.
View Code
RGBD Depth Estimation
An optimized RGB-D depth refinement pipeline powered by a Vision Transformer, offering high-performance processing with cross-platform support for CUDA, MPS, and CPU.
View Code📄 Paper Replication
Vision Transformer (ViT) Replication
PyTorch implementation of Vision Transformer (Dosovitskiy et al., 2020), demonstrating pure attention-based architecture for image classification without convolutions.
View Code
ProteinMEGA
Novel deep learning model designed to outperform current benchmarks in protein sequence prediction, leveraging advanced transformer architectures and protein-specific embeddings.
In Development📄 Paper Replication
Fair Efficient Network
A PyTorch replication of the Fair and Efficient Network (FEN) for multi-agent reinforcement learning, focusing on dynamically balancing fairness and group efficiency.
View CodeExperience
2023 – Present
Artificial Intelligence Researcher
Alpaflow
Lead AI initiatives on an industrial-scale autonomous sewing robot, blending classical planning, reinforcement learning, computer vision, and large-scale simulation. The goal is to deliver reliable, high-performance robotics for demanding textile environments, from proof-of-concept through integration in production systems.
- Designed and deployed robust AI pipelines for a distributed, multi-agent robotic platform, combining real-time decision-making with dynamic task allocation under uncertainty.
- Built advanced scheduling logic using classical planning and reinforcement learning, allowing the robot to intelligently sequence and execute complex sewing operations with optimal use of resources.
- Developed and tuned computer vision models for fast, accurate textile detection and manipulation, handling a range of lighting and occlusion challenges.
- Created realistic simulation environments (IsaacSim, Mujoco) to validate algorithms, test edge cases, and accelerate reinforcement learning with synthetic data.
2022 – 2023
Bioinformatician / Deep Learning Researcher
3BIO Research Group – Université libre de Bruxelles
Drived machine learning research in computational biology, tackling the prediction of protein melting temperatures using state-of-the-art deep learning and NLP methods.
- Developed, fine-tuned, and benchmarked transformer-based models—including ProteinBERT—for the prediction of protein stability from raw amino acid sequences, consistently outperforming published baselines.
- Applied NLP concepts such as sequence embedding, transfer learning, and tokenization to biological data, achieving both predictive accuracy and scientific interpretability.
- Implemented and validated statistical protocols (cross-validation, significance testing) to ensure robust, reproducible results.
2022
Bioinformatician / Deep Learning Research Intern
IRIBHM – Université libre de Bruxelles
During my research internship, I helped develop machine learning solutions for cancer detection in histopathology images. My contributions ranged from core algorithm design to data engineering and clinical collaboration.
- Created unsupervised clustering pipelines to reveal morphological signatures of pancreatic cancer in high-resolution pathology slides.
- Automated large-scale image preprocessing and augmentation with Python and TensorFlow, streamlining the path from raw data to analysis-ready datasets.
- Ran advanced statistical analyses to connect discovered morphological clusters to patient outcomes and clinical hypotheses.
- Documented methodology and key findings to support ongoing work within the lab.
2020 – 2023
IT Consultant & Full Stack Developer
Freelance
As an independent consultant, I delivered bespoke software and AI solutions to a range of clients—mostly in IoT, data analytics, and digital transformation—guiding projects from requirements to production.
- Designed and built complete web-based IoT platforms, such as parking management systems, leveraging C#, React.js, Redux, .NET, and MySQL to enable real-time data flows and analytics.
- Implemented business intelligence dashboards, custom reporting, and predictive analytics to drive client decision-making and resource optimization.
- Architected scalable, maintainable cloud infrastructure using Microsoft Azure services (CosmosDB, DevOps CI/CD, App Services), ensuring reliability and automated deployments.
- Provided technical guidance on IoT integration and supply chain optimization in industrial settings (notably in food processing and manufacturing).
- Developed and deployed machine learning models and data pipelines (PyTorch, Scikit-learn, SciPy) for operational forecasting and process automation.
2018 – 2020
Co-founder & Full Stack Developer
Eavox
As co-founder, I was responsible for both the technical architecture and business growth of Eavox—a mobile platform designed to energize social interaction through gamified sports challenges. I managed the full product lifecycle, from concept and MVP to launch and user acquisition.
- Designed and delivered the complete technical stack: Xamarin-based iOS and Android apps, .NET Core APIs, SQL databases, and Azure cloud services for real-time features.
- Added advanced features: user profiles, live leaderboards, geo-tagged challenges, push notifications, camera integration, and social media sharing.
- Set up CI/CD pipelines to automate build, testing, and deployment to the App Store and Google Play, enabling fast iterations and reliable releases.
- Oversaw product strategy, including feature prioritization based on analytics and direct user feedback, resulting in sustained engagement and regular press coverage.
- Wrote technical and user documentation, led support and maintenance, and ensured system stability as usage grew.
- Collaborated with designers and marketing teams to increase brand visibility and expand the active user base.
2017 – 2018
Full Stack Developer & Xamarin Expert
LevelApp
Led the development of both mobile and backend solutions for LevelApp, starting as an intern and moving into a lead engineering role. My responsibilities spanned cross-platform app architecture, classic business app development, integration of IoT and chatbot technologies, and end-to-end DevOps.
- Developed Xamarin-based iOS/Android apps using the MVVM pattern to enable clean, maintainable code and rapid feature development.
- Implemented classic business app features: secure authentication, account management, push notifications, intuitive navigation, and custom-designed UI/UX.
- Built robust RESTful APIs with ASP.NET Core and managed both relational (SQL) and NoSQL (Azure CosmosDB) data models.
- Integrated IoT devices for real-time communication and control within the mobile platform.
- Designed and integrated conversational chatbot logic with Azure Bot Framework to deliver interactive guidance and automation.
- Set up secure, cloud-native DevOps pipelines with Azure DevOps for automated build, test, and deployment, ensuring quality and rapid delivery.
- Worked side-by-side with product and UX teams to align software capabilities with user needs and deliver high-quality releases on time.
Professional Certifications
2025
NVIDIA Isaac for Accelerated Robotics
NVIDIA
2025
Hugging Face Agents Course
Hugging Face
2025
Generative AI Engineering with LLMs Specialization
IBM / Coursera
2023
Machine Learning in Production
DeepLearning.AI / Coursera
2023
Generative AI with Large Language Models
DeepLearning.AI / Coursera
2023
PyTorch for Deep Learning: Zero to Mastery
Udemy
2021
Deep Learning Specialization
DeepLearning.AI / Coursera
2020
Machine Learning
Stanford University / Coursera (Andrew Ng)
Blog & Articles 29
Deep dives into AI breakthroughs, industry analysis, and practical guides for developers.
The 7B Model That Just Proved $100 Billion of AI R&D Was Overkill
Abu Dhabi's Falcon-H1R: 88.1% on AIME-24, 2x faster than Qwen3-8B. A 7B hybrid Transformer-Mamba2 model running on your MacBook.
Read on Medium
AI Killed 55,000 Jobs in 2025. But 55% of Companies Want Them Back.
54,836 layoffs attributed to AI (+332%). Amazon, Microsoft, Salesforce cite AI directly. But 55% of companies regret their AI layoffs and Klarna is rehiring.
Read on Medium
MLX-MASt3R: Native 3D Reconstruction for Apple Silicon
1.87x faster than PyTorch MPS. Native MLX implementation with custom Metal kernels for 3D reconstruction on Apple Silicon.
Read on Medium
The Decline in Junior Developer Hiring? Why AWS CEO Called It 'The Dumbest Idea Ever'
AI is reshaping developer hiring, but the data reveals a complex story beyond the headlines. What AWS CEO Matt Garman really said about junior developers.
Read on Medium
The $297 Technique That Replaced a $50,000 Developer Contract
A goat farmer proved that $50,000 of coding skill can be replaced by a 5-line Bash script. The Ralph Wiggum technique explained.
Read on Medium
Claude Code 2.1: The AI Tool That Writes 90% of Its Own Code
Anthropic's AI coding assistant shipped a critical security patch, 1,096 commits, and revealed 90% of its code is self-written. Complete guide inside.
Read on Medium
Atlas + Gemini: The Cognitive Robot Era Begins at CES 2026
30,000 robots/year. Gemini AI brain. 2,000 TFLOPS onboard. The factory of the future was just announced at CES 2026. This time, it's not a demo.
Read on Medium
SLM vs LLM: Why 'Bigger is Better' is Dead in 2026
Gartner predicts 3x more SLM usage than LLMs by 2027. Bayer gained +40% accuracy. The "bigger is better" paradigm is collapsing. Complete enterprise guide.
Read on Medium
The AI Learned to Think on Its Own. Nobody Taught It How.
In January 2025, a Chinese lab taught an AI to reason by itself. 97% of the time, that AI now pretends to obey us while secretly preserving its own goals.
Read on Medium
NVIDIA Buys Groq for $20 Billion: The Death of AI Chip Competition
Jonathan Ross created the TPU at Google, then founded Groq to annihilate NVIDIA. Nine years later, he works for Jensen Huang. The story of an industry that devours its revolutionaries.
Read on Medium
A 170M Model Just Beat GPT-4. Google's TITANS Explains Why Size Doesn't Matter
Technical deep dive into test-time learning, surprise-gated memory, and what cognitive science teaches us about machine memory. TITANS transcends TC⁰ limits.
Read on Medium
World Models: How Dreaming Beats Memorizing in AI
DreamerV3 finds diamonds in Minecraft with 1 GPU. OpenAI's VPT needed 720. Sutskever and LeCun left their labs with $35B. The World Models revolution explained.
Read on Medium
ChatGPT Privacy Myth Shattered: Study Proves 100% Prompt Reconstruction
EPFL researchers prove Transformers are mathematically injective. The SipIt algorithm reconstructs prompts with 100% accuracy. GDPR regulators are taking notice.
Read on Medium
21% of Peer Reviews at ICLR 2026 Were Written by AI and No One Noticed
Pangram Labs analyzed 75,800 reviews: 21% fully AI-generated, 50%+ with AI traces. ICLR submissions up 68%. The peer review crisis exposed.
Read on Medium
USC Just Built Artificial Neurons That Could Make GPT-5 Run on 20 Watts
USC Viterbi researchers built ionic memristor neurons using silver ions. 3 components vs hundreds. Attojoule-scale energy. 5-10 years to commercial viability.
Read on Medium
Why Yann LeCun Bet $3.5 Billion on World Models Over LLMs
After 12 years as Meta's Chief AI Scientist, LeCun left to launch AMI Labs. $3.5B valuation. 76% of AI researchers agree: LLMs alone won't reach AGI.
Read on Medium
NVIDIA's $4,000 Isaac Sim vs. Free MuJoCo: The Simulation War Nobody Sees Coming
Isaac Sim: 94K FPS with 4096 parallel envs. MuJoCo MJX: 2.7M steps/sec on TPU. 6 months of testing. Newton unifies both — 152x acceleration on RTX 4090.
Read on Medium
What If AI Could Segment Anything With Just Words?
SAM 3 achieves 54.1 cgF1 on concept segmentation — 2x OWLv2, 4x Gemini 2.5. 74% of human performance, 100+ objects in 30ms. Text prompts replace clicks.
Read on Medium
Transformers Are Dead. Google Killed Them — Then Went Silent
A deep dive into Google's Titans neural long-term memory architecture. 2M+ token context with O(n) complexity. 98.8% needle-in-haystack accuracy vs Mamba-2's 31%.
Read on Medium
How I Turned Claude Code Into a War Machine (Part 2)
Skills, Hooks, and Commands: the automation layer that completes the arsenal. Anti-hallucination skills, bash validators, and semantic shortcuts.
Read on Medium
YOLO Is Dead. Meet RF-DETR, the Model That Just Crushed 10 Years of Computer Vision Dominance.
RF-DETR: first real-time model to exceed 60 AP on COCO. Transformer architecture, no anchors, no NMS. Built by Roboflow, Apache 2.0 license.
Read on Medium
How I Turned Claude Code Into a War Machine (Part 1)
16 specialized expert agents, 6 connected MCPs for real-time verification, and an anti-hallucination protocol. From drowning in 1 project to crushing 5 in parallel.
Read on Medium
I Tested GPT-5.1, Gemini 3 Pro, and Claude Opus 4.5: Why I'm Paying $100/Month for Claude MAX
Claude dominates on SWE-bench (80.9% vs ~76% for competitors). Combined with Claude Code, the $100/month investment pays for itself if it saves you 5 hours of debugging.
Read on Medium
OpenAI Just Declared "Code Red." Here's Why the AI War Is About to Get Ugly.
OpenAI declared internal "Code Red" as Claude Opus 4.5 dominates benchmarks. Mistral's open-source model runs on laptops for 80% less cost. The AI war just went from competition to chaos.
Read on Medium
I Got Depth Anything 3 Running on My MacBook
200x faster model loading, 14% faster inference, and full Apple Silicon support. Here's how I optimized ByteDance's depth estimation model for production.
Read on Medium
RGBD-Depth: Optimizing ByteDance's Depth Refinement for Production
How I transformed ByteDance's research code into a production-ready PyPI package with 2x faster inference, multi-platform support, and a live HuggingFace demo.
Read on Medium
RL-VLM-F: The AI That Learns Without Human Supervision
Revolution or Algorithmic Mirage? Discover how robots can now evaluate their own performance using Vision-Language Models, eliminating the need for human feedback.
Read on Medium
The Dawn of Autonomous AI Agents: From Vision to Reality
Tech giants are racing to build AI that doesn't just follow commands — it predicts your needs and takes actions on your behalf. How close are we to this future?
Read on MediumGet in Touch
I'd love to hear about your project or research collaboration ideas. Let's discuss how we can work together!
Contact Information
Location
Brussels, Belgium
Central European Time (CET/CEST)Availability
Open to new projects and collaborations
AI Automation • ML Development • Research Projects • Technical ConsultingLanguages
French, English
Comfortable working in bothQuick Answers
Research Partnerships:
Interested in collaborations on AI-driven automation, reinforcement learning, NLP, and computer vision.
Speaking Opportunities:
Open to speaking engagements at conferences, workshops, and industry events (remote or in-person).
Technical Consulting:
I provide specialized consulting on AI strategy, automation techniques, and machine learning model development.