Kuba Rogut is a deployed engineer at turbopuffer and creator of SFX Engine, an AI sound-effects platform used by thousands. His work spans creative AI products and the search infrastructure that helps coding agents retrieve relevant information from large repositories.
Rogut graduated in computer engineering from the University of Waterloo in 2024. Earlier roles included full-stack development at Scispot and Sobol, frontend engineering at Black Hawk Digital, cloud infrastructure with the AWS Hyperplane team, and financial software at Slash Financial. He became a founder in residence at Afore Capital in early 2025 and joined turbopuffer in February 2026.
He built SFX Engine for filmmakers, game developers, podcasters, and other creators seeking custom AI-generated audio. His writing on AI-assisted sound design emphasizes automating repetitive tasks while preserving human creative judgment. His other experiments include an AI-powered browser guide for navigating software, alongside storytelling, gaming, and financial applications.
- Embeddings as cached computation. Indexing a codebase requires upfront work, but creates reusable semantic context that multiple agents can query without repeatedly searching and rereading identical files. The savings become more significant across shared repositories and repeated sessions.
- TurboGrep and agent retrieval benchmarks. Rogut helped build TurboGrep, a command-line tool that chunks source code, generates Voyage Code 3 embeddings, and indexes them in turbopuffer. His Claude Code benchmark compared an unmodified agent, 50-line reading windows, and windowed reads supplemented with semantic search. File precision increased from approximately 65% to 87% in the semantic-search configuration, although recall varied by task.
- Hybrid agentic retrieval. Semantic search located behaviorally related code without shared keywords; grep performed better when following imports and explicit identifiers. Rogut favors combining vector similarity, full-text search, grep, and iterative reasoning, while recognizing that agents must learn when each tool is useful. Inline documentation can improve retrieval by clarifying what code does, and semantic indexing also supports audio, images, and video that cannot be searched meaningfully through filenames alone.
