Train Your First LLM on a Free TPU
Train Your First LLM on a Free TPU Here’s a fact that still surprises people:...
Keshan Sanjaya Sodimana
Jul 18, 2026
5 min read
keshan@blog:~$ ./home.sh
perspectives
In Search of Fellow Travellers
In Search of Fellow Travellers Disclaimer: This article was written by me, with AI assistance used only for grammar, spelling checks, phrasing and language improvements. All views and experiences shared...
./read_more.sh~/articles
Explore my thoughts on Machine Learning, AI, and more
Keshan Sanjaya Sodimana
The Ratchet Loop — Systematically Optimizing a TPU Kernel to the Hardware Ceiling
The Ratchet Loop — Systematically Optimizing a TPU Kernel to the Hardware Ceiling Part 3...
Keshan Sanjaya Sodimana
Apr 28, 2026
19 min read
Profiling TPU Kernels - XProf, HLO, and the Roofline Model
Deep-Dive into TPU Profiling — XProf, HLO, and the Roofline Model Part 2 of the...
Keshan Sanjaya Sodimana
Apr 25, 2026
26 min read
Implementing a Fast Attention Fusion Kernel
Writing Fast Attention on TPU — From Naive Kernel to Fused FlashAttention with Pallas Part...
Keshan Sanjaya Sodimana
Apr 21, 2026
27 min read
Running Containerised Batch Jobs on Google Cloud Platform
Running Containerised Batch Jobs on Google Cloud Platform I needed to OCR tens of thousands...
Keshan Sanjaya Sodimana
Apr 19, 2026
14 min read
From Video to Structured Insight: A Practical Guide to Gemini-Powered Video Analysis
In today’s data-driven world, video content is everywhere—from security surveillance and industrial monitoring to social...
Keshan Sanjaya Sodimana
Jun 21, 2025
8 min read
The Art of Controlled Randomness: A Deep Dive into Sampling techniques in LLMs
Large Language Models (LLMs) are powerful predictors. Given a sequence of text, they excel at...
Keshan Sanjaya Sodimana
Apr 18, 2025
14 min read