About Muhammad Ahmad — the human runtime

I design and ship agentic AI systems with enforceable trust boundaries — and I've organized my whole life around making that claim checkable. This page is the receipt for the person behind the console.

01Path, compressed

I'm 18 and this is on purpose: Chiniot, Punjab; school through Intermediate in Computer Science (completed 2026); university interrupted by circumstances that weren't optional, which made "employable by evidence" the only viable plan. I had roughly two focused hours a day — most people underestimate what 730 honest hours compounds into. The visible result: 2024–2025 doing sales and client web work to pay for the attempt, April 2026 shipping my first public repos, June onward building xr, late July a from-scratch transformer, August it all hardened into CI, releases and audits. Three dated steps beat one adjective.

02How I build

I use AI coding agents the way a senior engineer does: I write the contracts, review every line, and let tests decide. The xr history shows the pace, so I'll say it plainly — much of it is orchestration, not typing. What I own: the architecture, the policy gate, the budget governor, the tests that fail when the code lies, and this site's habit of showing receipts. Ask me about any line and I can defend all of it, including the AI-written parts, because I read them before they shipped.

That's not a disclaimer; it's the actual 2026 skill, and the only version of the story that survives an afternoon of clicking through commit history. I'd rather build the process where a lie is detected nightly than the prose where it's merely unlikely.

03What's next — live list, not a vision board

• First external PRs (3–5 merged). The one gap every audit of me has found: 65 PRs, all into my own repos. Targets: promptfoo's eval API, LangGraph, Ollama. When I ship one, this sentence dies and gets replaced with a link.
• A real evaluation pipeline. Golden datasets + LLM-as-judge scoring with significance tests — the evals muscle my portfolio should prove next, not claim next.
• AegisRoute rebuild. One-evening repos stay in /lab until they pass the same CI bar as xr. Breaking the start-fast-finish-late habit is a feature of my process, and this is where it gets proved.

04Context I won't pretend isn't relevant

I run RRRTX Systems, a one-person product studio (custom e-commerce + AI automation) with real delivered client work — the reason I care about costs, failure modes and maintenance is that a client's invoice is how I learned all three. Before code, two years of sales taught me the harder half of engineering: finding out what someone actually needs. The Optimized Human newsletter keeps the philosophy muscle warm; The Lost Signal in /lab keeps the fiction one. Everything I ship is public, including the embarrassments — a hackathon project once marketed cryptography its code didn't have; that lesson is why every sentence on this domain is clickable proof.

05Tools, honest fields only

Daily: TypeScript + Bun + Neovim inside zsh (yes, checked); Python when the job's a script and PyTorch when it's science; Vercel, Cloud Run, Docker, GitHub Actions; Ollama for everything that should never phone home. Editor loyalty is to the test runner, not the mode line.

resume: request via email — I'd rather ship one that matches this site 1:1 than leave a stale PDF linked.

snapshot: 9 tracked repos · newest push yesterday · say hi