Research Scientist at Microsoft · ML Systems

Thamme “TG” Gowda

I build AI systems that survive the trip from research to production.

My work spans multilingual and multimodal models, data and distributed training, evaluation, quantization, GPU kernels, and low-latency inference.

Selected work View CV Kidi demo (experimental)

Email tgowdan [at] gmail [dot] com

Featured work

Research that ships.

Selected outcomes across production AI, model inference, algorithms, and systems software.

Local inference · September 2026

Kidi: lightweight native AI inference

I built a C++23 inference toolkit for CPU and Apple Metal, with an experimental WebAssembly port. It runs Gemma text generation and RTG translation without PyTorch or Transformers at inference time.

Native quick start and code Experimental browser demo Architecture and requirements

The experimental browser demo is my first attempt at WebAssembly; it's still slow at the moment. CPU-only, with no WebGPU acceleration. Prompts stay on your device. About 2.49 GB of model files, cached after download; current 64-bit Chromium and ample memory recommended.

Production ML · 2026

Offline translation in Microsoft Edge

Quantized and integrated translation models for private, on-device translation through the built-in Translator API.

Announcement API

Tahoma W4A8 inference pipeline using asymmetric 4-bit weights and dynamic FP8 activations

Research leadership & systems · 2025–2026

WMT Model Compression and TahomaMT

I lead the WMT Model Compression shared task and built its two-year evaluation infrastructure for runnable systems, controlled H100 execution, and quality–footprint–speed trade-offs. The 2026 edition received 41 participant-submitted variants from 13 teams. TahomaMT tests those ideas in a custom C++/CUDA runtime.

Findings Shared task Evaluation harness TahomaMT paper

80× over Python gzip C++ · Python · Go · Rust · WASM

Systems software · 2026

pigzpp: compression across the stack

A thread-safe C++23 compression core with multi-language bindings, native ZIP, and a fast PNG codec.

Code Paper Benchmarks

per merge

~~O(N)~~ O(log N)

max-heap + lazy deletion

Algorithms · 2026

Faster BPE training in SentencePiece

Replaced repeated linear scans with a heap-based update and merged the implementation upstream into Google SentencePiece.

Upstream PR Write-up

Earlier systems

Built to outlast the experiment.

Research infrastructure that became useful beyond the work that started it.

Multilingual data and training

MTData, RTG, NLCodec, and Many-to-English

I built an open toolchain for reproducible MT data, vocabularies, training, and inference, then used it to train a single model spanning more than 500 source languages. As a WMT General MT organizer from 2022 through 2026, I have maintained versioned MTData recipes that give participants reproducible constrained-track datasets across five editions.

MTData WMT26 WMT25 WMT24 WMT23 WMT22 RTG NLCodec Paper

Open model deployment

NLLB Serve

A practical web interface, REST API, and batch decoder for deploying Meta's No Language Left Behind models across 200 languages. It remains my most widely adopted personal repository.

Code PyPI

Distributed systems

Sparkler

I created Sparkler at USC: an extensible web crawler built around Apache Spark, Kafka, Solr/Lucene, Tika, and distributed JavaScript rendering. I later handed the project to its maintainers when I shifted focus to my Ph.D.

Code

Research

Selected publications

All publications

2026
Findings of the WMT 2026 Shared Task on Model Compression: No Free Lunch at Extreme Compression
Forthcoming at the Eleventh Conference on Machine Translation (WMT 2026)
Paper Project Code
2024
PyMarian: Fast Neural Machine Translation and Evaluation in Python
Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations
Paper DOI Project
2020
Finding the Optimal Vocabulary Size for Neural Machine Translation
Findings of the Association for Computational Linguistics: EMNLP 2020
Paper DOI Code
2018
Deep Mars: CNN Classification of Mars Imagery for the PDS Imaging Atlas
Proceedings of the AAAI Conference on Artificial Intelligence
Paper

Notes

Recent writing

All posts

Aug 2026 Is 4-Bit Quantization Truly Lossless?
Quality accounting, W4A8 kernels, and one-H100 inference results.
Jul 2026 Making Compression Faster
One accelerated core across data pipelines, Docker, PNG, ZIP, and the browser.
Mar 2026 From O(N) to O(log N): Faster BPE Training
The algorithm, implementation, benchmarks, and upstream journey.

Background

Research depth, engineering range.

At Microsoft, I work on multilingual and multimodal AI and the systems that train and serve these models. I have also helped organize the WMT General MT shared task for five consecutive editions, from 2022 through 2026. Previously, I spent five years as a Research Engineer at USC ISI and worked as a Data Scientist intern at NASA JPL.

I earned my Ph.D. and M.S. in Computer Science from USC. My doctoral work studied rare phenomena learning in neural machine translation (dissertation). I have also built software at startups, co-founded Datoin, and served as an Apache Software Foundation committer and PMC member.

Invited talks

Plenary speaker at IEEE ICKECS 2026

Delivered Efficient Software for AI Systems: From Training to Inference at the 4th IEEE International Conference on Knowledge Engineering and Communication Systems, hosted by SJC Institute of Technology on April 24, 2026. Slides Program

Invited speaker at SJCIT FDP 2026

Presented LLMs and Multilingual AI: Enabling Vernacular Innovations in India for faculty and graduate students in SJCIT's Faculty Development Program on January 22, 2026. Slides

2022–Now: MicrosoftSenior Research Scientist
2017–2022: USC ISIResearch Engineer
2016–2017: NASA JPLData Scientist intern
2022: USCPh.D. in Computer Science

Beyond the work

I grew up in a farming family in southern India, where work was measured in acres and yield, not hours. That upbringing still shapes how I approach hard problems: understand the whole system, stay practical, and judge the work by what it produces.