I build software where research meets real systems: model runtimes, data and training infrastructure, performance-critical libraries, and tools that make research easier to reproduce and deploy.
This is the part of the work I was able to open source. The known unknowns live, rather appropriately, in private repositories.
Current Systems
WMT Model Compression: Benchmarking Deployable Systems
I lead the WMT Model Compression shared task, now in its second edition, and build the open evaluation infrastructure behind it. The WMT25 harness standardized offline Docker submissions; WMT26 advances to self-contained runnable systems with common install and inference contracts, sanity checks, reproducibility recipes, organizer-run execution on controlled H100 hardware, and Pareto analysis across translation quality, model footprint, GPU memory, and decoding speed.
The second edition received 41 participant-submitted variants from 13 teams. Its forthcoming findings report shows that precision alone does not determine efficiency: quantization becomes fast only when the serving runtime is designed to exploit it, and quality must remain part of the comparison.
Tahoma: Low-Bit AI Systems in C++
Tahoma is a C++23 runtime and research platform for training and inference across language, multimodal, and classification models. Current work includes low-bit quantization, custom CUDA and ROCm kernels, paged attention, continuous batching, CUDA graphs, multi-GPU execution over NCCL/RCCL, and faster rollouts for reinforcement learning. The WMT26 evaluation tests TahomaMT’s FP8 and INT4 paths against diverse participant systems, grounding kernel and runtime work in translation quality rather than throughput alone.
pigzpp: Fast Compression Across the Stack
A clean-room C++23 rewrite of pigz that became a reusable compression core for C++, Python, Go, Rust, and WebAssembly. It supports standard gzip/zlib streams, native ZIP containers, and fast PNG encoding over zlib-ng and Intel ISA-L backends.
PyMarian: Marian Inference and Evaluation from Python
Python bindings to Marian NMT, connecting its optimized C++ engine and CPU/CUDA backends to Python workflows. PyMarian supports translation, COMET-style evaluation, notebooks, and prebuilt model applications while retaining the performance and memory advantages of Marian.
Multilingual AI Infrastructure
MTData: Reproducible Machine Translation Data
MTData automates locating, downloading, parsing, caching, and citing parallel corpora. Its versioned recipes make datasets reproducible across experiments.
As a WMT General MT organizer from 2022 through 2026, I have maintained the official MTData setup for five consecutive editions: version-pinned recipe files and command-line workflows that let participants reconstruct the constrained-track training data. The published coverage grew from 11 WMT22 recipe IDs to 21 WMT26 recipes; later editions added parallel caching, compressed materialization, corpus statistics, and quality-estimation workflows.
NLLB Serve: Multilingual Translation as a Service
A web interface, REST API, and batch decoder for deploying Meta’s No Language Left Behind models across 200 languages. It packages model loading, language handling, GPU execution, and interactive or programmatic translation behind a small interface.
RTG and NLCodec: Training, Inference, and Vocabularies
RTG is a PyTorch-based neural machine translation toolkit, and NLCodec provides inspectable word, character, class, and BPE codecs with Python, CLI, and PySpark interfaces. Together with MTData, they formed the toolchain behind a many-to-English model spanning more than 500 source languages.
SotaStream: Streaming Training Data
A streaming data pipeline for large-scale machine translation training. SotaStream builds composable generator graphs for on-the-fly mixing, sampling, filtering, and augmentation, avoiding rigid preprocessing pipelines and unnecessary materialization of every experiment variant.
Earlier Foundations
Sparkler: Distributed Web Crawling and Content Analysis
I created Sparkler at USC and designed its core architecture: an extensible web crawler built around Apache Spark, Kafka, Solr/Lucene, Apache Tika, and distributed JavaScript rendering. It combined scalable crawling, fault tolerance, near-real-time indexing, content analysis, and a plugin system. I later handed maintenance to the project team when I shifted focus to my Ph.D.
More on GitHub
These are selected projects, not an exhaustive catalog. Smaller experiments, research artifacts, teaching material, and earlier systems are available in my complete list of public GitHub repositories.
I also participate in Stack Overflow Q&A.