Machine learning engineer and DevRel working on model inference: making models fast and helping developers use them. Also into philosophy and art, part-time fashion model.
At Superlinked, as a founding member of technical staff, here are some public contributions:
- TopK-Embed-V1 in SIE - I added two multi-vector embedding models (0.8B and 2B) to SIE v0.9.0, Superlinked's open-source inference engine. Fused GPU kernels, CUDA graphs and padding-free batches make them 1.5 to 1.6× faster than the vendor's pipeline. One query takes 6 ms instead of 81 ms. The top search results stay the same. I also fixed a bug that made full batches of large outputs fail in the cluster.
- SIE - SIE runs small models for agents. I contributed to the project and also led developer growth and adoption and more than doubled its GitHub stars in five months.
I co-authored a paper on FlashNorm.FlashNorm merges the RMSNorm weights into the next linear layer, so transformers run faster.
- Open source contributions in transformer-tricks, I merged 10+ pull requests. They add GPU benchmarks, Gemma 4 support, and a guide to use FlashNorm with your own model. With FlashNorm, Llama-3.2-1B runs 12.77% faster in Hugging Face Transformers.
- Weight Folding, CUDA Streams, and the Bug That Made My Model Speak Backwards - AI Engineer World's Fair 2026 · video
- The Small Model Infrastructure Nobody Built (So We Did) - AI Engineer Europe 2026 · video
- One GPU, Four Retrieval Modes: Multi-Model Search Serving - Berlin Buzzwords 2026 · video
- From BM25 to Mixture-of-Encoders - Haystack EU 2025 · video
- What Actually Makes Embedding Model Inference Fast? - article
- langchain-superlinked - PyPI package with a custom Superlinked mixture-of-encoders retriever for LangChain
- Superlinked x LlamaIndex - A custom LlamaIndex retriever that uses Superlinked
- AdalFlow - Contributor to the library for building and optimizing LLM task pipelines: multimodal OpenAI support, the integrations page, and RAG and text-splitter tutorials
- Mood-based song recommender - Transformer models and vector search to recommend Spotify songs by mood (Qdrant Vector Space Talks)
- Data science and ML roles in retail, fintech and biomedical AI