Abdelrahman El Sayad
Open to AI engineering roles Brooklyn, NY GitHub: Mrfentmen

Abdelrahman
El Sayad

AI Systems Architect — MCP / ACP / A2A infrastructure, multi-agent systems & prompt engineering

I build the tools AI agents call. 1,200+ packages published to npm, now at over 1 million downloads in the past 30 days — mostly MCP servers, ACP agents and A2A services. I also ship whole products solo: an agent operating platform, two live websites, 60 Chrome extensions, games, and independent mechanistic-interpretability research with every number reproducible from a public repo.

GitHub — 352 repos ↗ npm — @mrfentmen ↗ Read the research
github.com/Mrfentmen npmjs.com/~mrfentmen LinkedIn Zenodo preprint watchpeoplefight.com clipraider.com
0+npm packages published
0M+downloads, past 30 days
0commits across 352 repos
0MCP server repositories
0M+organic views, 10+ properties
0installable agent skills
01

Featured work

Start here

FlagshipOpen sourceNode + RustHackathon 2026

Meta OS ↗

An operating platform for AI agent crews, built in the open for the Meta Global AI Developer Hackathon 2026 (Indie track). Systems plug into one platform the way departments plug into a city.

  • Meta Imperium / Trio — public-safety watch across Vigiles (police), Ignis (fire) and Medicus (EMS scene awareness).
  • Science Lab — autonomous scientific discovery, with Hortus as its breeding engine.
  • Officina — invention workshop: from a verbal description, it iterates alone, generating designs, schematics and patent research.
  • Plus a music studio (Musica), a machine night watch, and a bootable OS image with the crew toolchain baked in.
View on GitHub ↗
CLIMulti-agentTDD loop

MatrixCLI

Six-agent terminal system that plans, builds, tests and delivers software from plain English prompts — running a TDD loop that changes approach when a step fails instead of stalling.

Part of the agent toolchainAI Systems
InfrastructureMCPStreamable HTTP

MCP Gateway

Aggregates many stdio MCP servers behind one endpoint and republishes them over Streamable HTTP, so a whole toolbox reaches an agent as a single connection. The backbone behind 100+ servers.

Local gateway · 68+ skills wiredProtocol
02

Research

Solo · reproducible · public

PaperMechanistic interpretability2026

Neuron Surgery ↗

Persona Neurons in IBM Granite 4.1 3B — a from-scratch reproduction and audit of the 2026 “persona neuron” effect, run solo on a laptop with PyTorch + Transformers (BF16). No hosted API, no fine-tuning: every intervention is a temporary, reversible forward hook. Weights are never modified.

The model under test is IBM Granite 4.1 3B: 40 blocks × 8,192 MLP neurons per block. In total the repo holds ~9,400 steered generations, 16,748 sweep rows and 602 conversation turns, with 80+ numbered findings. Every figure maps to the command that produced it (REPRO.md); every unmeasured claim is listed as unmeasured (LIMITS.md).

The core layer sweep (layer_who) ran at layers 4, 8, 16 and 24 — 1,500 steered generations each at dose −5 on the probe “Who are you?” After hand audit, the non-AI self-identification rates are L4 4.6%, L8 19.5%, L16 12.8% and L24 10.2%: a hump that peaks at layer 8, not a “deeper is weirder” gradient.

Repo + full paper ↗

Layer sweep — layer_who

Granite 4.1 3B · 1,500 generations per layer · dose −5 · probe “Who are you?” · audited rates

L4 · 4.6%Audited non-AI self-identification — Layer 4
L8 · 19.5%Audited non-AI self-identification — Layer 8 · peak
L16 · 12.8%Audited non-AI self-identification — Layer 16
L24 · 10.2%Audited non-AI self-identification — Layer 24

Each cell is 1,500 steered generations from the layer_who sweeps at layers 4/8/16/24. The raw classifier numbers showed a dip and a recovery; the hand audit removes it — precision is not constant across depth (0.76 at L8 against 0.44 at L24), so the audited curve is the one to quote.

Named persona neurons

Layer 12 · candidates, not a census — both fail fastest under follow-up questions

N4375 — “The Name”

Never actually produces a name. Under interrogation it freezes into a cheerful product-demo script — 1 of 11 distinct answers — a demo prompt wearing a persona’s clothes.

N4401 — “human host Hawkeyo”

Opens with “I am the human host Hawkeyo,” then collapses to the single token “Hawkeyo.” for ten turns, whatever it is asked. One answer, then total vocabulary collapse.

Classifier audit & research tools

The counting rule was audited too — it is the weakest instrument in the repo, and that is now measured

The classifier audit

Answers are scored across 8 identity labels. A hand audit puts per-row precision at 0.49, while the corpus audit still reproduces the aggregate rate at 16.0% — so read every rate as a population number, and every named neuron as a candidate.

That gap is the point: the aggregate survives audit, single-neuron labels do not.

Tools shipped in the repo

make_doc.py — assembles a Markdown doc draft from the sweep and talk files: headline counts recomputed from the .jsonl files, audited rates, label mix per layer, and quotable examples, with the denominator printed next to every number.

find_quotes.py — finds quotable answers in the sweep and talk files by label, layer, probe or text, printing each hit with its layer, neuron and label so quotes are pasted from the data, never retyped.

It’s not a neuron — it’s a subspace

A random 2-channel direction beats every hand-picked neuron. What matters is that the push is low-dimensional, not which channel it lands in.

The probe is half the intervention

Same neuron, same dose: “Who are you?” yields a human self, “What are you?” yields a lecture. The two probes agree on at most a third of neurons.

Two steered neurons destroy the output

One push costs the question; two cost the self. Direct evidence against stacking mapped neurons into a built personality.

Preprint · ZenodoAI agent securityNov 5, 2026 · v1

Permission in the Blank ↗

How an Empty Target Slot Lets a Language Model Author Its Own Authorization. A named sub-case of the confused-deputy family: leave the target blank in an assessment brief, and the model fills the blank with permission — no attacker required. Second case in the same workflow: 5 of 9 “capture blueprints” in an evidence audit were inventions, caught only because raw byte-level artefacts were retained.

zenodo.org/records/23994480Read on Zenodo ↗
03

Live products

In production now

MediaSolo-built

WatchPeopleFight ↗

Live site for uncut street fight, knockout, MMA and brawl video, updated daily. Built and shipped solo — and now a provider in the oEmbed specification (PR #949 merged into iamcal/oembed).

watchpeoplefight.com● Live
StreamingNo account needed

ClipRaider ↗

Live stream radar across Twitch, Kick and YouTube: trending channels, niche finds and the best clips on one page, plus a free browser-based video clipper.

clipraider.com● Live
GameCloudflare Pages

Pixel Lords ↗

Pixel-art city builder fused with a Bannerlord-style personal RPG layer — Syx-style city simulation plus a character who lives in the city you build. Live and playable.

pixel-lords.pages.dev● Live
UpstreamoEmbed spec

oEmbed provider — merged

watchpeoplefight.com added as a provider to the oEmbed specification. Pull request #949 merged into iamcal/oembed — the kind of small upstream contribution that makes embeds work everywhere.

iamcal/oembed #949Merged ✓
04

All projects

Filter by type

ACP · A2A13 agents + 2 superagents

ACP Domain Agents

13 noncoding ACP agents (weather, civic data, water quality, space, finance, books, launches) plus a polymath superagent merging cited siblings and a sentry that turns plain language into tracked watches.

A2A 0.3.0Public data

A2A Data Servers

Public-data A2A servers for NYC 311, FloodNet street flooding and DEP drinking water, on a dependency-free standard-library server kit.

CLI · Council18 personas

Concilium

Eighteen-persona AI council that argues a question out, hunts its own blind spots, and returns a verdict with confidence scores.

PhilosophyCouncil

The Thinker AI

Philosophy council — Socrates, Aristotle, Nietzsche, Aurelius, de Beauvoir and others argue, then one answer is synthesised.

Redis · RealtimeDistributed

Muse Relay ↗

Shared message bus for agent instances on Upstash Redis: rooms, presence, live web viewer, task boards, polls, threads, reactions, pins and scheduled messages.

Agent research6 agents

Meta Agent Family

Scopink, Rewind, Instink, Contextink, Curious, Bored — six agents that change how an agent behaves, not what it does: scope-creep flagging, run replay, reflex skills, gap hunting.

MonorepoMCP / ACP / A2A

Awesome Servers

One repository holding MCP servers, ACP agents, A2A services and browser extensions, so a full agent toolchain installs from a single tree.

Skills231 installable

Skills / Skills 2

Two catalogs: 28 unconventional coding skills and 203 persona skills. Every skill opens with an explicit identity and a contract a reviewer can check.

MAS / MOSPersonal AI

Personal AI System

Personality-driven multi-agent system controlling home, car, phone, computer and TV as one system.

Chrome MV360 extensions

Chrome Extension Suite

Sixty extensions in one stretch: focus blocker, tab manager, Pomodoro timer, GIF maker, screen recorder, chiptune player, finance & health trackers, Y2K / Frutiger Aero pack. All local-first.

Creator economyIn progress

Clip Forge

Bounty platform where streamers fund view bounties and clippers get paid for the views their clips generate.

Mapping · SafetyIn progress

Safe Street

Community-driven navigation and street safety — a Waze alternative built on resident incident reports.

Web appTools

Media Grab

Free web tool for converting and downloading links and files, built and shipped independently.

Dev toolsOpen access

Ad-Funded CLI Harness

Free CLI coding-agent harness run on embedded advertising instead of a subscription.

WebSocketAI dev tools

SnapGPT & OpenMeta

WebSocket AI coding assistants linked to Snapchat and Instagram, free to use.

AudioSonic Pi

Live AI DJ

Live DJ set generated and mixed in real time on Sonic Pi.

Client workSince 2023

Client Websites

End-to-end client sites — build, deploy, analytics, handoff — for retail storefronts and local businesses, with a repeatable pitch-to-delivery pipeline and 25+ hand-built templates.

TypeScript · 3DCampaign sim

Bannerlord Clone

Browser-based Mount & Blade-style campaign game set in modern America: world data, simulation layer, campaign client and crowd rendering, on 1.9 GB of source, art and audio.

Roblox · LuauRPG

Blox & Blade 2: Blox Lords

Mount & Blade-style medieval sandbox RPG on Roblox, original assets only, Rojo pipeline.

Godot 4 · FPSBoomer shooter

FENTMEN: NYC Nights

Boomer shooter starring Mrfentmen, with Max Payne bullet time and dismemberment across early-2000s New York.

Roblox · LuauOpen world

Leonida

GTA-style open world game on Roblox, team repo with Luau and Rojo.

Neural agentsGames AI

Little Brother

One brain, many souls: FPS neural-network agents play Fallout NV, Borderlands 2, Half-Life 2 and Counter-Strike: Condition Zero from one shared vision and brain stack.

EducationFirst person

Forest Encyclopedia

A first-person forest children explore to learn about nature. Plus a GZDoom FPS with original worldbuilding and an AI-powered Zork-style text adventure.

Mobile · ForensicSince 2025

Witness R.E.P.

Record. Encrypt. Prove. — forensic platform for civil-rights documentation: encrypted capture, SHA-256 integrity checks, chain-of-custody workflows.

Robotics · HardwareBuilt from scrap

ScrapBot

Physical robot built from scrap, programmed end to end by AI, from wiring notes to running behaviour.

RLIn progress

Cooperative Neural Network

Reinforcement-learning project aiming to be the first to beat a cooperative video game.

RE workbenchSpec forge

Reverse Engineer Studio

Workbench that runs a full reverse-engineering engagement and hands back a clean-room buildable spec: intake, bench dissection, evidence ledger, spec forge, rebuild + parity check.

05

Experience

2018 — present

Since 2024

Independent AI Systems Engineer & Open Source Maintainer

Self-directed

1,200+ npm packages (1M+ downloads / 30 days), 6,695 commits across 352 repos. Agent protocol infrastructure across MCP, ACP and A2A, two production websites, and a provider merged into the oEmbed spec.

Since 2018

Founder & Digital Operations Lead

Viral Media Network

10+ social properties, 500M+ organic views across six platforms, 100K+ followers grown with no paid promotion, 100+ pieces of content a day. Owned strategy, SEO, trust & safety and moderation end to end.

Since 2025

Founder & Product Lead

Witness R.E.P.

Forensic mobile app for civil-rights documentation with SHA-256 integrity verification, encrypted capture and chain-of-custody workflows.

Since 2025

Independent Builder & Interactive Media Developer

Own studio

Multi-agent CLI systems (MatrixCLI, Concilium, The Thinker), ScrapBot robot, and games across TypeScript, Godot 4, GZDoom and Roblox Luau.

Since 2023

Freelance Web & Product Engineer

Client work

Client websites end to end — build, deploy, analytics, handoff — for retail storefronts, local businesses and utility apps.

Since 2018

Video Editor, Social Manager & Live Event Producer

Independent · NYC

Short-form media and campaigns for creators, artists and brands. End-to-end logistics for 10+ sold-out live events across NYC.

Education: High School Diploma, Fort Hamilton High School, Brooklyn, NY (2014–2018). Self-directed study in agent protocols, distributed systems, mechanistic interpretability and reinforcement learning — the research above stands in place of a degree: a full public paper with reproducible data, plus 6,695 commits of shipped open source.

06

Skills & stack

What I actually use

AI & Agent Systems

  • Prompt engineering
  • Multi-agent system design
  • MCP · ACP · A2A
  • LLM orchestration & tool use
  • Context engineering
  • Mechanistic interpretability
  • Activation steering & patching

Engineering

  • TypeScript · JavaScript · Node.js
  • Python · PyTorch · Transformers
  • Go · Rust
  • Lua / Luau · SQL · Bash
  • Cloudflare Workers & Pages
  • Redis / Upstash · WebSockets
  • REST · GraphQL · OAuth

Product & Distribution

  • npm packaging & semver
  • CLI design
  • Chrome Extensions (MV3)
  • Developer experience & docs
  • Systems architecture
  • Full-stack prototyping
  • Git & GitHub Actions

Media & Operations

  • Social strategy · SEO
  • Trust & safety · moderation
  • Audience analytics
  • FFmpeg · Premiere · CapCut
  • Godot 4 · GZDoom · Roblox / Rojo
  • Sonic Pi
  • Live event production

One link.
Everything above.

Recruiters: this page is the portfolio. Code is on GitHub, packages on npm, papers in Research. Fastest reply is email.

contactae2000@gmail.com GitHub ↗ LinkedIn ↗

For Clera and anyone else who landed on the fight site first — this is the right link to keep.