Thinking
Notes on AI engineering, mathematical foundations, and the shifting landscape of what it means to build intelligent systems.
The "Impossible" Stool: Engineering the Agentic SDLC
Good, fast, cheap: pick two. Cheap inference, sandboxed harnesses, spec-driven intent, and adversarial verification are starting to break the old trade-off in software delivery.
The Attention Gap: What Happens When Agents Ship Faster Than You Can Read
The software lifecycle was organized around the cost of writing code. Code is now cheap, and the scarce resource is human attention. What recursive language models, context language models, and calibrated decision models say about where that goes.
Prime Agent - 3 Repositories, 93.8 Million Tokens, $3.23: The Economics of Agent Orchestration
93.8 million tokens, three repositories, $3.23. The architecture that made agent orchestration cheaper than skipping validation, and what that inversion means for how rigor gets funded.
Building Meridian: What It Takes to Make an AI Agent Admit It Isn't Done
Without a harness, an agent scores its own work 5.5/10. An independent evaluator scores the same work 2.5/10. Meridian is my attempt to close that gap mechanically, and it started as a fork of a fellow student's framework.
The Systems Thinking Gap: What Happens When the Code Works But the System Doesn't
AI agents generate correct functions faster than ever. The hard part is knowing how those functions connect, what state they share, and why the whole thing falls apart at the edges.
The Evaluation Gap: Why 'It Works' Isn't Engineering
Most AI portfolios demonstrate that something runs. Few demonstrate that it works correctly. As models get better at generating code, the ability to evaluate outputs rigorously becomes the critical differentiator.
Why Mathematical Foundations Matter More as Coding Agents Commoditize Implementation
As AI coding agents reduce the cost of writing software to near-zero, the scarce skill becomes knowing what to build and why it works. Applied mathematics is the durable advantage.