Research on quality control in crowdsourcing, consensus protocol design for proof-of-stake networks, tooling for AI coding agents, and the software and datasets that came out of all three.
Protocols & technical reports
A validator-selection algorithm that weights blockchain validators by earned, context-dependent reputation rather than stake alone. Designed on Cosmos SDK/Tendermint.
A quantitative look at how decentralized delegated proof-of-stake networks actually are.
Publications
Software & datasets
Source-available forensics for AI coding-agent spend. Analyzes Claude Code sessions for usage patterns, errors, and token costs (late compaction, oversized tool results, redundant re-reads), free for personal and internal use. Source on GitHub.
A Claude Code skill that makes coding agents do the UX work before writing UI code: an enforced process with sourced rules, validated by scoring with-skill and without-skill builds of the same prompt against a 30-check audit. MIT licensed.
A crowdsourcing evaluation dataset: 300K+ contributions from 400 workers over 1,000+ questions and 500+ tasks, with declarative worker profiles and self-evaluations across knowledge domains.
CReate, Enrich, eXtend: a framework for building and extending crowdsourcing datasets, with clustering-based task selection and campaign-site generation.
Learns offline which worker profiles are reliable for which task types, then selects workers online from the available crowd.
Talks
Conference talks at HASE 2016 (Orlando), COMPSAC 2017 (Turin), and ADBIS 2019 (Bled), and a series of IRIXYS workshop talks (2014–2018, France/Germany/Italy) on quality control in crowd computing.