Jonathan Esperanza
Software engineer
I work on Unifies signals from Mail, Search, Finance, Shopping, and Scout into versioned, agent-ready profiles for 80M users. Extraction runs at 2,000 events a second, and agents retrieve memories with hybrid keyword and semantic search. and A self-serve fine-tuning and RL platform built on vLLM, LoRA, and GRPO. Its first model replaced Gemini 2.5 Flash Lite for memory extraction with 2.5x lower latency, 2.1x lower cost, and more faithful output (0.92 vs 0.79). at Yahoo. I like building AI that’s practical and effective, shaped by my research on the side: My one-person lab exploring the hard parts of building software in the agentic era: source control for hundreds of agents changing one repo every day, a fast isolated workspace for every agent, and one record of who changed what across every agent vendor. In private development., long-horizon agents, and RL.
Outside of that, I enjoy Memento, the Ocean’s series, and The Prestige. and FC Barcelona..
Previously, I worked on Recommendations Platform, promoted from SWE I to SWE IV. Ran the pipelines behind 1.5B notifications a month, moved model serving to Triton, and cut $2.6M a year in data costs. at Credit Karma.
Writing
First post coming soon. Subscribe with RSS to get it.
Projects
- rust-build-bench How much CI time Rust monorepo build tools save on real pull requests
-
Open source Contributions to Apache Beam, Trino, and Bento