Thamme “TG” Gowda
AI Research Scientist • LLMs • Multilingual & Multimodal AI • High-Performance ML Systems
Email: tgowdan [at] gmail [dot] com
·
Curriculum Vitae
Thamme “TG” Gowda is a Senior Research Scientist at Microsoft, where he works on large language models, multilingual and multimodal AI, and high-performance ML systems. He earned his Ph.D. (2022) and M.S. (2017) in Computer Science from USC. Previously, he was a Research Engineer at USC ISI (2017–22; for five years!), a Data Scientist intern at NASA JPL (2016–17), a Committer and PMC member at the Apache Software Foundation, a full-stack software developer at a startup (2012–14), and co-founder at Datoin.com (2014–15).
His work spans the full model lifecycle — data pipelines, large-scale pretraining, post-training alignment (SFT, RL), evaluation, distributed training, quantization, and low-latency inference — bridging ML research and hardware-aware systems engineering. Selected highlights:
- Quantized and integrated translation models for offline translation in Microsoft Edge.
- Building multimodal LLMs for production applications.
- Enhanced training-data quality to improve production model quality across hundreds of languages.
- Built a 500+ language many-to-English translation model.
- Tinkering with Tahoma, a personal C++23 playground for learning GPU systems (custom CUDA/ROCm kernels) and experimenting with my ideas for AI.
Earlier, his Ph.D. research addressed the rare phenomena learning problem in neural machine translation (dissertation, slides).
News
- 2026-07 — Co-organizing the WMT26 Model Compression and General Machine Translation shared tasks (@ EMNLP 2026, Budapest).
- 2026-06 — On-device, offline translation shipped in Microsoft Edge (145+ languages) via the built-in Translator API — see the Edge announcement.
- 2026-06 — New post: A Field Guide for Sanity-Checking Your Training Pipeline.
- 2026-04 — Fast BPE learning merged into Google SentencePiece: O(N)→O(log N) per merge, with 10x–20x+ speedups in large settings (note on X).
- 2026-03 — New post: From O(N) to O(log N): A Faster BPE Training Algorithm.
- 2026-03 — New post: I Let Two AI Agents Race to Modernize pigz.
Education
2018–2022
2015–2017
2008–2012
Professional Career
2022/11 – Now
2017/07 – 2022/10
2016 – 2017
2015 – 2016
2014 – 2015
2012 – 2014
More?
I speak Kannada/ಕನ್ನಡ, Telugu/తెలుగు, and Hindi/हिन्दी (+English, of course!). I was raised by a family of farmers in a southern Indian village. So, beside the "professional" experience that I have displayed here, I also have much more experience in agriculture that you will not hear: vineyard, silkworms, dairy, sheep, poultry…!
I came to this grad school for learning the artificial intelligence, but I also ended up learning a bit about myself/consciousness:
While coding an artificial soul for machines
He wins his gifts, defeats his curses
And discovers his own soul that thinks and feels!
Consciousness is divine [Prajnanam Brahman]
This self/Atma is divine [Ayam Atma Brahman]
You are that [Tat Vam Asi]
I am (part-of) it [Aham Brahmasmi]