Notes from the machine room.
I'm Karthik Rajkumar. I write long-form about how machine learning works, the systems that carry it at scale, and the agents we are now building on top of both — each in its own publication, each with its own desk. The research behind it is published formally, in journals and patent filings.
The masthead
Choose your desk
Four publications with one editorial standard: explain the idea properly, show the trade-offs, and leave you able to build it yourself.
Machine Learning
The Meridian
Models, maths and the craft of making them work.
Long-form notes on machine learning — from the intuition behind a loss function to what it takes to ship a model and keep it honest in production.
System Design
The Blueprint
How large systems are drawn, built and kept standing.
Architecture teardowns and design essays — queues, caches, consistency, and the trade-offs that decide whether a system scales or stalls.
Agent Architecture
The Dispatch
Field notes on building AI agents that actually finish the job.
Patterns for planning, tool use, memory and evaluation — the architecture behind agents that are reliable enough to trust with real work.
Research & Patents
The Proceedings
Peer-reviewed papers, patents and the ideas behind them.
The formal record — journal papers and patent filings, each with its abstract, citation and a plain-language account of what it contributes.
Fresh off the press
Latest stories
- 1
A tensor is a shape, some strides, and a pointer
Every machine learning library is built on one data structure. Here is what a tensor actually is, why it is mostly a description rather than data, and how that one idea makes transposing a matrix cost nothing.
17 October 2026·9 min - 2
Warnings as errors, and a Miri job with nothing to check
Before Ferro had any real code, every check was switched on and set to fail the build — including a tool that hunts for bugs in code the project does not contain yet. An empty codebase is the only time that is painless.
28 September 2026·6 min - 3
Automation in the language you're already using
Ferro's project automation is a Rust program, not a Makefile or a folder of shell scripts. Why the xtask pattern pays for itself, and the small details that make automation pleasant to live with.
28 September 2026·6 min - 4
Enforceable Trust for Enterprise AI Agents
Enterprises grant AI agents trust once and keep it forever. This paper makes trust a computed, decaying score that decides — at the moment an agent asks for credentials — how much it is allowed to do.
29 September 2026·5 min - 5
GraphZero Agent: Monte Carlo Graph Search as Reasoning-as-a-Service
Today's agents reason along a single path and rarely look back. This paper adapts AlphaZero-style search to general agentic reasoning — over a graph rather than a tree — and offers it as a service any agent framework can call.
14 March 2026·3 min
“Understand it deeply enough to draw it on a whiteboard.”
The house rule