Salvatore Sanfilippo

Sanfilippo created Redis in 2009 and led the project until stepping back in 2020. He returned to the company in 2024 as an evangelist, describing a role that connects the community with the company through examples, documentation and ideas for the database. His account of returning emphasizes programming and design alongside learning from how people use the software.[2]

A data structure for similar items

His Vector Sets work brings vector similarity into Redis through a set-like interface. Elements can be added, removed and queried for neighbors without exposing every implementation detail to the application. Sanfilippo explains modifications to an HNSW graph that support actual node deletion and preserve connectivity through reciprocal links and relinking. The contribution combines a compact public interface with changes to the underlying search structure.[4]

A narrow engine for local inference

DwarfStar specializes in selected model families rather than trying to run every possible model. It combines native inference paths, model-specific quantization and hardware support including Metal, CUDA and ROCm. SSD streaming and distribution across machines address memory limits. Its README credits llama.cpp and GGML as essential references and identifies retained or adapted components, making the project’s relationship to that prior engineering explicit.[3]

In his writing on AI-assisted programming, Sanfilippo argues that a developer’s attention should move toward design ideas, behavior and thorough verification as code generation becomes faster. He describes questioning the model’s implementation choices and comparing results with other systems, rather than treating a generated program as complete because it runs. The argument grows from his work on DwarfStar’s inference correctness and performance.[5]