Jules Belveze
Research Engineer · AI Agents
I build agents that can improve safely.
At H, I research self-evolving computer-use agents and how to evaluate changes without regressing what already works.
Previously, I built agentic software and community at Dust, worked on anomaly detection at Microsoft, and scaled multilingual NLP at Hypefactors.
Writing
-
Harnesses Are Becoming State
-
Evaluation First: The Hidden Prerequisite for Self-Evolving Agents
-
LayerSkip: Early Exiting Grows up for LLMs
-
BERxiT: Early Exiting Beyond Entropy
-
Instruction Fine-Tuning Evaluation and Advanced Techniques (opens in a new tab)
-
DeeBERT: Teaching BERT When to Stop Thinking
-
Early Exiting: The Under-Hyped Compression Method
-
Instruction Fine-Tuning Fundamentals (opens in a new tab)
-
Scaling Machine Learning Experiments With neptune.ai and Kubernetes (opens in a new tab)
-
Case Study: MLOps for NLP-powered Media Intelligence using Metaflow (opens in a new tab)
-
Scaling-up PyTorch inference: Serving billions of daily NLP inferences with ONNX Runtime (opens in a new tab)
-
Atlastic Reputation AI: Four Years of Advancing and Applying a SOTA NLP Classifier (opens in a new tab)