Now
Only what is true right now.
Working on
Building products and features that combine domain rules and AI for meetings, customer support, finance, and business analysis at KRAFTON
Improving data validation, enterprise integrations, operational recovery, and reporting in business-owned products
Can RL reward and verification signals be trusted: research in progress
What verifiers of ML agents' improvement claims miss, and why time-series anomaly-detection metrics disagree: research in progress
Building
jevcompat: run on eight Jev-compatible servers, two passed, and maintainers of four shipped fixes within three days
sessionreel: long Claude Code sessions as 30 to 60 s videos, on PyPI
An open-source toolkit for time-series anomaly-detection evaluation
Reading
Whether the reward and verification signals used to train AI agents with RL can be trusted
Evaluation design and statistical inference
Operating and verifying agent systems