Writing
Notes on AI, security & building things that last
Occasional essays on local-first AI, resilience and compliance, and the craft of building software people can actually trust.
- June 22, 2026
The AI race just fractured — and the US did it to itself
The US banned its own best models, Japan matched frontier benchmarks without a single frontier model, and China’s open-weight stack is closing fast. The AI landscape from two weeks ago is already gone.
Read the post - June 13, 2026
AI just became a single point of failure — the fix is local
The U.S. government switched off the most powerful AI model on Earth overnight. Here’s why that’s the strongest argument yet for running AI you actually own.
Read the post
From LinkedIn
More short-form posts
I share most of my day-to-day thinking on LinkedIn — from AI and cybersecurity to heritage and seva. Here are a few recent ones.
AI threats now outpace human defenders
AI can bring down a company faster than a human team can even detect the breach. On the “automation threshold,” why compliance isn’t security, and why threat modeling has to turn into immediate engineering action. (BSides Vancouver)
Launching SikhArchive V2.0 — 500 years of literature, open to all
The SikhLibrarian project found a permanent home at sikharchive.net: thousands of digitized texts from 50+ scholars, a 758-million-word corpus, restored with AI and free for the world.
AI beat Apple’s M5 chip security in 5 days
Apple shipped the M5 with billion-dollar hardware defenses; a research team using AI found a way in within five days. Why “wait and see” is now a dangerous security strategy.
Cybersecurity as a growth accelerator in FinTech
Security gets dismissed as a cost center until a crisis strikes. Why secure-by-default infrastructure is actually a competitive advantage in financial services.
Week 2: the two hardest problems building the AI Sikh Librarian
OCR managed only ~40% accuracy on printed Gurmukhi and failed entirely on handwritten manuscripts — so I built GurmukhiFix. Plus: earning the trust of eight Sikh organizations through the right license.
An AI librarian for 500 years of Sikh literature
A fully local RAG pipeline over 758 million words across five languages, with Research and Learn modes and Chicago-style citations — making manuscripts modern OCR couldn’t read finally searchable.