ML Researcher & Builder
Bartosz Dzionek
I do machine learning research at Exa, where we’re building search that’s actually made for AI. Previously at Google. I care about the full stack — from the research that makes a model better to the engineering that gets it in front of people.
My interests
AI Search for AI
Building ML models for search that is designed for other ML models. I dream about the retrieval layer that lets LLMs reason about any public or private information.
Multimodality and multilinguality
Expanding LLM research breakthroughs to search over data in any modality and language. Building representations allowing cross-modal retrieval.
Research → production
I like owning the whole arc: from proof of concept to the system that ships it to real users. I build agents to speed up this process.
Selected writing
All posts →- Hello, world (and a feature tour) A starter post that shows off the things this blog supports — LaTeX math, syntax-highlighted code, and embedded images.
- Search built for models, not people Web search was designed for humans skimming ten blue links. AI systems want something different — and that gap is the interesting problem.
- The low-rank softmax bottleneck, explained Why a language model’s final softmax layer can quietly cap how expressive its predictions are — and what you can do about it.
Selected projects
State-of-the-art reranker
I co-authored a SOTA reranker with top NDCG@5 on BEIR and >2x better latency than any competitor.
Research · LLMLow-rank softmax bottleneck
Undergraduate thesis on the low-rank bottleneck in LLMs. I surveyed existing solutions and showed they can also help in image transformers.
Research · LLMLLM for rap generation
Fine-tuned GPT on 20k rap songs to produce a 2x better rap generator as measured by perplexity.
Engineering · AI4ScienceCamera trap classifier
Built an end-to-end solution for camera-trap research to simplify wild animal classification. Saved zoologists many hours with a good precision-recall model.