Who is Muhammad Ahmad?

An agentic AI engineer based in Chiniot, Pakistan. Ships the xr agent runtime and a trained from-scratch foundation model, does client work through his studio RRRTX Systems, and publishes every claim with a verifiable source. Completed ICS (Computer Science) in 2026; no degree, never claimed one.

A TypeScript/Bun agent runtime with a spend ceiling agents cannot overrun, fail-closed policy gates, and an offline-verifiable audit chain. MIT-licensed, published on npm as @rrrtx/xr, with 2,924 tests passing and a release audit that also documents its CI failures.

Is Muhammad Ahmad available for hire?

Yes — open to remote junior/associate agentic-AI or AI-platform engineering roles and contract work on agent systems, automation, or evaluation pipelines. Email is the fastest path; replies within a day, timezone-wise.

Does Muhammad Ahmad use AI coding agents?

Yes, openly: he writes the contracts, reviews every line, and lets tests decide. The full policy is stated on the site's about page under "how I build" — the position is that AI-assisted work is judged by the same audit bar as hand-written work.

fetched yesterday

Muhammad Ahmad

I build agentic AI systems with enforceable trust boundaries.

Runtimes, evals and audits — spend ceilings agents can't overrun, policies that fail closed, logs you can verify offline. Everything on this site links to its proof; CI re-checks those links nightly and admits it when one rots.

view the work github contact

HOW I BUILD AI-assisted development, human-verified: I write the contracts, review every line, tests decide. the manifesto →

◉ 31.72°N 72.98°E — Chiniot, Punjab · remote-first UTC+5 (Karachi)

claims → evidence · sample

full ledger: run `evidence` — every row opens something you can check yourself
claim	status	evidence
The agent loop cannot overrun a token or dollar ceiling — enforced in code and covered by tests	TESTED	test/cost.test.ts ↗
Publicly installable from npm as @rrrtx/xr	PUBLISHED	npmjs.com/package/@rrrtx/xr ↗
Forty-nine-finding forensic self-audit run before release; every finding closed with an evidence note	AUDITED	docs/audit/FINAL_AUDIT.md ↗

> ls ~/featured

01Live · active

xr

A local-first AI agent runtime that cannot overspend its budget or hide its actions.

721commits 65merged PRs 7×CI workflows v3.1.5on npm

TypeScript · Bun · SQLite · MCP

case study repo live demo

02Active

xr-foundation-model

A tiny GPT grown from zero — tokenizer, training loop, math proofs — and then audited by me, in public.

✔ 49-finding forensic self-audit, closed 49→0
✔ val perplexity 301.55 — measured, then admitted it's small
✔ checkpoint committed to the repo as evidence (76.8 MB)
✔ 33 test files incl. ground-truth math proofs

Python · PyTorch · from-scratch BPE · pytest

case study repo

commit-canvas

Live

case study repo live example

Turn any git repository into an animated, shareable story page. Zero cost, zero auth, one command. Starred by three early users, its own Actions pipeline renders this repo's story page on every push — dogfooding included.

Python · HTML/JS · GitHub Actions · 0 network requests

04 Gemma MonitorLocal-LLM news intelligence running in production for $7.40/month — the paid API bill it replaced was $22. Shipped 05 CivicShield AIA civic assistant that would rather refuse than hallucinate: retrieval before generation, fail-closed. Shipped

> view all work

real output · every line linked to its source

No fake logs, no invented telemetry. This is a replay of what my own release audit and CI actually recorded — failures included, because that's the point of keeping them visible.

ahmad@xr:~% — session replay · all output real · typed once per visit

$ bun test
238 files · 2924 pass · 13 skip · 0 fail · 12,978 expects · 57.6 s
$ curl -fsSL https://raw.githubusercontent.com/ahmadrrrtx/xr/main/install.sh | bash
binary installed · @rrrtx/xr v3.1.5 · MIT · golden path 0.76 s e2e
$ gh run view 31702859699   # cross-platform CI — as my own audit documents it
✘ linux / macos / windows failed at step 7 “Full unit suite”
  real, environment-specific failures (100 s / 93 s / 326 s) — not timeouts, not hidden
$ cat docs/audits/XR_FINAL_RELEASE_AUDIT.md | grep -i verdict
✔ “No evidence of fabricated test success or vacuous coverage was found.”

every line above, sourced: test + CI verdicts ↗ baseline audit json ↗ live CI ↗

watch · repo + verifier snapshot baked at build — refreshes on every push

git log -1 xr@ea21bd4 "Merge pull request #75 from ahmadrrrtx/phase2/decisi…" · yesterday

fetched yesterday

the full history, warts and velocity ↗

role	Agentic AI Engineer
focus	runtimes · evals · audits
status	● open to work
location	Pakistan · UTC+5
latest	@rrrtx/xr v3.1.5
links checked	nightly →

host Muhammad-Ahmad · Chiniot, PK

shells zsh, bash (both real)

editor neovim + AI agents — disclosed

uptime 2 focused hours/day, ~7d/wk

focus Agents / LLMs / Systems

claims 0 unverified on this site

2026-09First external open-source PRs — evaluating promptfoo's eval API while writing my own golden-dataset harness.
2026-09xr: decision-boundary hardening phases merging on branches; the audit-chain spec under review.
2026-09AegisRoute rebuild: same idea, xr-grade CI discipline. One-evening repos are the pattern I'm breaking.

the honest version of this list →

I Ran Gemma 4 on a $7/Month Server and Built an AI-Powered News Monitor That Costs $0 to OperateWhat I bring to engineering: measured cost tables, real failure handling, no hand-waving.
The Agent That Writes Its Own Manual: Hermes Agent's Self-Improving ArchitectureHow I read someone else's system: architecture-first, framed as analysis, not authorship.

> cat ~/writing

email → ahmadrrrtx@gmail.com github → github.com/ahmadrrrtx · client work → rrrtx-systems.com writing → ahmadrrrtx.substack.com ↗ — newsletter, occasional engineering notes, unsubscribe anytime also on · substack ↗· medium ↗· hashnode ↗· dev.to ↗· indie hackers ↗· daily.dev ↗

open a conversation