Ryan Codrai is a Member of Technical Staff at Anthropic and the developer of TurboVec, an in-process vector index implemented in Rust with Python bindings. His research includes evaluating synthetic brain MRI images.[1][3]
Start with TurboVec →
Career
Contributions
Search directly over compressed vectors
TurboVec applies Google Research’s TurboQuant approach to a local vector index. It rotates vectors and quantizes coordinates with a codebook, then uses SIMD lookup tables to score queries against the compressed representation. Incremental additions and deletions do not require retraining a codebook. Its Python and Rust interfaces expose search, filtering and persistence.[2]
Measuring the quality of generated images
With Matthias Treder and Kamen Tsvetanov, Codrai co-authored a study comparing human judgments and automated measures of GAN-generated brain MRI images. The work finds that familiar image-quality metrics can track large differences while a learned quality model is more sensitive to subtle differences among better images. It recommends combining several kinds of analysis rather than relying on one score.[3]
Ideas
The repository publishes benchmark scripts and describes the limits of its design. TurboVec scans the corpus, so query cost grows with the number of vectors. Quantization is lossy, and recall should be checked on the intended data. It is a library rather than a hosted service, with no built-in replication or network API.[2]
Tools, writing & talks
Tools
Start here
TurboVec
github.comA Rust and Python index that searches compressed vectors locally and supports incremental updates.
Read the Python example and the section on when to use another index. Trying it requires Python or Rust and representative embeddings; compare recall on your own data before relying on the compressed results.
Creator and maintainer[1][2]Original source
Writing, research & talks
Research comparing expert ratings with distortion metrics and learned image-quality assessment.
Co-author[3]