I publish at top AI venues — and turn the results into products.

Multi-Agent Orchestration for Terminology-Constrained Machine Translation in Industrial Localization

A deployed four-agent pipeline — analysis, translation, post-editing, review — with glossary injection. 99.4% terminology accuracy and the highest ChrF2++ across three language pairs, beating all 20 systems on the WMT25 terminology task. pp. 917–926.

Cited as related work by researchers at JPMorgan AI Research (Findings of EMNLP 2026) and Amazon (WMT 2026).

AIDA Agents: A Multi-Agent Translation Platform with Context-Aware Quality Control

With Piotr Peszynski. Translator → arbiter/rater → post-editor → re-rater, optional RAG, MQM-inspired explainable scoring; best on 10/11 WMT24++ pairs and 70–98% publication-ready on industrial data.

Auditing LLMs’ Hidden Political Stances via First-Order Logic

With Rakshit Naidu. Frontier models are trained not to take political sides in plain English. Ask the same question in first-order logic and, across four contested topics, commitment to one side rises from 21.7% to 48.2% on average; one leading provider went from 32% to 85%. Neutrality is a property of the phrasing.