Jules Belveze
ML systems engineer · writer
Make machine learning earn its compute.
I work on ML systems where model quality, latency, and cost pull in different directions.
At Dust, I build product infrastructure. Earlier, I worked on anomaly detection at Microsoft and multilingual NLP at Hypefactors, including systems serving about a billion inferences a day.
- ~1Binferences per day
- NLPmodels in production
- MLefficiency research
Recent Posts
-
LayerSkip: Early Exiting Grows up for LLMs
Why do decoder-only language models still run every token through every layer?
-
BERxiT: Early Exiting Beyond Entropy
Why should BERT trust its own confidence scores to decide when to stop thinking?
-
Instruction Fine-Tuning Evaluation and Advanced Techniques (opens in a new tab)
You fine-tuned your model to follow instructions—but how do you actually know it works? This guide unpacks the evaluation frameworks and parameter-efficient methods that separate production-ready agents from expensive science projects.
-
DeeBERT: Teaching BERT When to Stop Thinking
Why does BERT need twelve layers to classify “I love this movie” as positive?
-
Early Exiting: The Under-Hyped Compression Method
Why are we burning GPU hours to answer “2 + 2 = 4”?