01
Architecture
Which agent, what context, what data, and where its authority starts and ends.
Getting an agent to do something is easy. Being able to trust what it does isn't.
I design the architecture and controls needed to take an agent into production.
I'm Alex. I've spent 10 years building software. The last four have been about agents: how to give them context, how to measure them, and how to stop a convincing output from being mistaken for a correct one.
Right now
Over the past year, I've been helping drive the shift to AI inside my company.
Not just testing tools: finding where they create value, defining how to use them, and helping teams adopt them without losing control of what they produce.
I work inside the real codebase, on what actually ships. Not from the outside with a slide deck.
akrcontext , open source, with a client running it in production.
Benchmarks and experiments published here, including the code and the failures.
What I work on
02
Controls and evaluation
How you verify that its output is correct. An agent reviewing its own work is not a guarantee.
03
Delegation and adoption
How much can be delegated, under what guarantees, and what needs to exist so it doesn't become expensive later.
Recent experiments
View all → AI Engineering
Angel-Devil: I Pitted a Judge Against an Attacker on the Same AI-Generated Code. The Attacker Won.
I ran a judge and a breaker against the same AI-generated code —same model, same spec. The breaker found bugs the judge explicitly approved as correct. Why adversarial feedback beats evaluation in coding agents.
AI Engineering
Does RLM Help Understand Large Codebases? I Measured It
I compared three paradigms —vanilla, terminal agent, and RLM— on a 1.13M-token codebase. RLM helps small models, but a terminal agent matched it at 3× lower cost.
AI Security
Prompt Injection in Your IDE: The Attack That Starts with git clone
I built four prompt injection demos and ran them against Claude Code, Codex, Gemini CLI, and Copilot. The results aren't what you'd expect: modern models catch the obvious attacks. The one they don't catch is the most dangerous.