AI that helps with real work
I care about the part after the demo. Does it help? Can people trust it? What breaks when real users arrive?
Hi, I’m Sai.
I love solving interesting problems with AI.
I dig into systems, try new tech, and build things to see what happens.
A few things I’m into right now
Things I keep coming back to.
The systems nobody sees
Schedulers, queues, latency, and the small decisions that make complicated software feel simple.
Building in health
Hard constraints lead to better questions about safety, trust, and what a product should do in the first place.
Recently
New articles and ideas.
Agent Evals Need Decisions, Not Essays: Building with Jev
I built a small Jev evaluator to see what changes when models return typed judgments and code owns the final pass, review, or fail decision.
Read the article →Delivering AI models globally is a systems problem
Training gets attention. Reliable delivery is where architecture becomes real.
Read the article →What scheduling algorithms teach us about serving LLMs
Fairness, throughput, and tail latency are old systems problems wearing new AI clothes.
Read the article →This should be a conversation
Come say hello.
I want this site to lead to better conversations, not just more page views.
- Tell me what you are building.
- Send me something worth reading.
- Disagree with one of my articles.