CTO, The Non-Functional Co. · AI compliance for regulated finance · LLM agent reliability & determinism
Focus areas: LLM agent reliability · context & memory management · extracting deterministic behaviour from non-deterministic models · compliance for regulated industries.
arXiv:2601.07190 (cs.AI) · Jan 2026 · sole author
Focus — an agent that autonomously compresses its own context, inspired by slime-mold exploration: 22% token reduction on SWE-bench Lite with no loss of accuracy.
Reflections on AI systems, engineering, and software development
Getting reliable, side-effecting tool use out of a small, cheap model (Haiku 3) by investing in the harness rather than the model — ref proxies that make hallucinated IDs impossible, and tool contracts that make skipping a required step impossible.
Token-per-minute limits refill gradually, so backing off past ~60 seconds is just wasted time. A three-layer pattern for rate limiting LLM APIs across distributed workers: shared pause state, jittered resume, and a single retry owner.
An in-container NGINX broke SSE streaming, corrupted auth responses, and blocked cluster mode. Why I pulled it out, what had to be rebuilt in-process, and where compression actually belongs.
Deeply nested JSON makes per-node jobs unaddressable and partial loads impossible. Flatten to one row per node with order and indent integers, then rebuild hierarchy with SQL range queries or a single O(n) pass.
AI Agents are stuck with a single chain of thought. If can be encouraged to traverse different paths, learn from them, and go back in time with the learnings. That would be cool.
LLMs often fabricate UUIDs in RAG pipelines. Learn two production-ready strategies - enum constraints and token aliasing - to cut hallucination rates and stabilize multi-turn reasoning.
I moved from a React to a Vue project two years ago from a company change. If you are about to...
You might already have some quick answers in mind for this question. I totally get it - I do...
TL;DR This is a long post detailing my journey of migrating a project from Vue 2 to 3. It...
XPath selectors offer powerful querying capabilities but face compatibility challenges with modern web technologies like Web Components and Vue 3.
How I converted PhantomCSS from a single-user tool into an integrated part of our team's workflow, covering challenges from CI integration to test execution optimization.
How we used the power of regular expressions to speed up our migration from a global namespace pattern to AMD modules (RequireJS), automating the conversion of 400+ files in Badoo's MobileWeb project.
Master five powerful debugging techniques that web developers can use to reduce debugging time and squash challenging bugs: Weinre, DOM Breakpoints, the debugger statement, hooking native methods, and mapping remote scripts locally.
Overcome the challenges of implementing fixed headers in mobile web apps by hiding the address bar, removing interaction delays, and preventing unwanted page scrolling when using form inputs.