Selected work
ParaCrawl & MaCoCu
Web-scale multilingual data pipelines >1 PB processed
I led engineering for large-scale pipelines that acquired, processed and curated multilingual web data for machine translation and language-model research.
My work covered crawling, Common Crawl and WARC processing, distributed execution, text extraction, language identification, alignment and data cleaning across HPC and cloud infrastructure with international research partners.
Focus: multilingual NLP web data distributed computing data quality
Stack: Python C++ Snakemake HPC and cloud Common Crawl WARC
Bitextor
Open-source pipeline for multilingual web data
I was a core developer and maintainer of Bitextor, an open-source system for harvesting, extracting, aligning and cleaning multilingual web content from a single machine to distributed computing environments.
The project has been used in large European language-data initiatives including ParaCrawl and MaCoCu and combines web processing, NLP, orchestration and high-performance computing.
I also co-developed Monotextor, its monolingual counterpart for harvesting and curating web corpora, covering WARC/web ingestion, text extraction, language identification, cleaning and deduplication.
Focus: crawling text extraction language identification alignment filtering
Stack: Python C++ Bash Snakemake distributed systems HPC
Reverso Context
Production NLP and search >7M daily queries
I contributed to the data and retrieval pipelines behind Reverso Context, improving search accuracy for multilingual examples drawn from aligned corpora.
The production NLP system served more than seven million queries per day as part of the wider Reverso product ecosystem.
Focus: multilingual NLP information retrieval parallel corpora production scale
Visit Reverso Context
MultiScore
Transformer-based music transcription
I led the engineering work for MultiScore, a research project developing Transformer-based models for optical and audio music transcription within a shared end-to-end framework.
The project brought machine learning research, evaluation and music-domain knowledge together to turn score images and musical audio into structured notation.
Focus: Transformers multimodal ML optical music recognition automatic music transcription
Visit the project
CrossPoint Reader
Systems and open-source engineering
CrossPoint is community-built, fully hackable open-source firmware for ESP32-C3-based Xteink X3 and X4 e-readers. My work as a project maintainer has included EPUB rendering, constrained-memory reliability, performance, testing and localization.
Focus: embedded C++ parsing constrained-memory systems testing
View on GitHub
CARAMEL & CRYPTOTRACK
Applied AI research and development
My work at Treelogic has included two AI initiatives: CARAMEL, exploring personalized cardiovascular risk assessment for women during the menopausal transition; and CRYPTOTRACK, monitoring cryptocurrency transactions to detect illicit activity.
