42.3314° N, 83.0458° W
LOCAL --:--:--
STATUS: OPEN TO EXCEPTIONAL TEAMS
Detroit, MI · Software systems
Self-taught engineer building AI agent systems, security tooling, and embedded hardware.
Systems that physically exist in the world — complete products, often shipped in 24 hours.
Four records, in order of weight. Each one names the institution, the number, and the proof class behind it.
[EV-01]
82 / 5,700+
All 7 tasks · top ~1.5%
One of 82 solvers from 5,700+ participants across 500+ US institutions. Solved all 7 tasks in reverse engineering, malware analysis, cryptography, and forensics.
[EV-02]
60s → 10s
80% latency reduction
Selected by Google Cloud to collaborate directly with their engineers on a production AI agent. Diagnosed and fixed a latency bottleneck through parallelized, distributed tool orchestration. The session was filmed and published as a public episode.
[EV-03]
~9,900
Confirmed unique recorders
Independent internet-scale Shodan research characterized Dahua-lineage surveillance device exposure across Morocco's IP space. Over half of the confirmed recorders were in CVE-flagged vulnerable families. Accepted as a poster at USENIX Security '26.
[EV-04]
5
Wins in under 6 months
Wins include MIT x Harvard, Duke's HackDuke first place in the Financial Track, Michigan State, University of Michigan-Dearborn's Best Use of Gemini API, and York College of Pennsylvania's grand prize, won solo.
Selections and roles beyond the four records above, each with its proof class.
30,000+
R-01Applicants
Selected from more than 30,000 applicants.
30 / 8,000+
R-02Selected students
One of 30 selected from more than 8,000 applicants; speaking at Adobe Education Summit.
2026–2027
R-03Full program
Accepted into the full 2026–2027 program at the only Apple Developer Academy location in North America.
Leader
R-04AWS Student Builder Group
Leads Detroit's AWS Student Builder Group, organizing cloud and AI events to grow the city's tech ecosystem.
Ambassador
R-05Microsoft student program
Represents Microsoft on campus, running AI and cloud workshops across Detroit's student community.
[NATIONAL PRESS · MOROCCO]Feature portrait · English edition · July 27, 2026Read at telquel.ma (opens in a new tab)
Covered by Morocco’s national press and television after the Google Cloud collaboration.
Medi1 TV
[NATIONAL TV]La Matinale · Héros du jour
Studio interview on his journey and why curiosity outweighed formal education
National morning-show feature introducing his work to a Moroccan audience.
Open source article (opens in a new tab)
La Nouvelle Tribune
[PRESS]Feature
“Meet Sami Maghnaoui, the 18-year-old Moroccan who caught Google's attention”
English-language feature on the same work.
Open source article (opens in a new tab)
CF-01Playback IQ
Google Cloud AI Agent Clinic
AI soccer analyst narrated live by a Gemini-based agent, with tactical breakdowns and previews.
60s → 10s · −80% latency
Built solo, then selected into Google Cloud's AI Agent Clinic. Working alongside Google engineers on the production bottleneck, Sami parallelized the sequential tool-call chain that was serializing the analysis — response time fell from 60 seconds to 10. The session was filmed and published as a public Google Cloud episode.
Problem
Generate live tactical analysis from a replay of Morocco vs. Portugal in the 2022 World Cup.
Build
Shipped the initial build solo, then worked with Google Cloud engineers on the production latency fix.
Architecture note
Parallelized distributed tool orchestration reduced response time from 60 seconds to 10 seconds.
CF-02IRIS
Yale
Voice-driven AI lab assistant on camera-enabled goggles.
<2s end to end · 3 AI models
One teaching assistant for 40 students means most lab mistakes go uncaught. IRIS watches each experiment step through a goggle-mounted camera and answers by voice: an ESP32-S3 streams video to a FastAPI backend chaining a vision model, Gemini 2.5 Flash, and ElevenLabs speech — the full loop lands in under two seconds. When Google deprecated the team's model mid-hackathon, the pipeline was swapped live without missing the deadline.
Problem
Verify lab steps visually while delivering instructions, answers, and spoken guidance through goggles.
Build
Built ESP32 firmware and a FastAPI backend orchestrating a vision model, Gemini 2.5 Flash, and a voice model end to end in under 2 seconds.
Architecture note
Structured per-step prompting made verification dramatically more reliable.
CF-03The Underlayer
MIT Reality Hack · AWE 2026
Spatial cybersecurity for Snap Spectacles.
BLE/WiFi → holographic AR
Two of four teammates quit mid-hackathon; the remaining two shipped anyway. A wearable ESP32 scanner streams nearby BLE/WiFi devices into a single FastAPI service that checks them against the OSV vulnerability database, and Snap Spectacles render each device as a hologram anchored where it physically sits. Suggested fixes execute over SSH only after explicit human approval — the system never acts alone.
Problem
Make an invisible RF attack surface spatially visible.
Build
Built a custom ESP32 BLE/WiFi scanner feeding a FastAPI backend that cross-references discoveries against the OSV vulnerability database and renders each device as a holographic AR card.
Architecture note
Scanner discoveries flow through vulnerability matching before rendering as spatial device cards.
CF-04Counterfeit Eye
University of Michigan-Dearborn
Real-time counterfeit banknote detection deployed to an iOS app.
1,200+ images · 24 hours
Marketplace scammers pass fake bills because no accessible tool exists to check them — and no public dataset existed to build one. The team shot 1,200+ photos of banknotes under varied lighting, fine-tuned a vision model on a $15 MacBook, and shipped it behind a Flask API to iOS; training finished five minutes before the deadline. On a positive detection, Gemini locates the nearest police department.
Problem
No suitable counterfeit banknote dataset existed.
Build
Shot a 1,200+ image dataset by hand, fine-tuned a pretrained computer vision model, and deployed it through a Flask API to an iOS app built in 24 hours.
Architecture note
The model finished training minutes before the deadline.
CF-05Counterfeit Mind
Michigan State · SpartaHack 11
Cryptographically signed NFC banknotes verified by AI vision, hardware, and blockchain.
4 verification layers · solo
A solo build rethinking banknote security from the manufacturing line up: each note carries an NFC chip signed with a Treasury Ed25519 key. A phone photo lets Gemini extract the note's visual fields, an ESP32 with an NFC reader verifies them against the signed chip data, and an Ethereum ledger catches revoked or duplicated serials — four independent defenses, each catching a different class of forgery.
Problem
Banknote security features are physical, static, and increasingly forgeable.
Build
Built a Swift iOS app, ESP32 firmware driving a PN532 NFC reader and TFT display, and a Node.js relay that verifies Ed25519 signatures and registers serials on Ethereum Sepolia.
Architecture note
Treasury-signed NFC data, visual field matching, signature verification, and on-chain registration each catch a different class of forgery.
CF-06BlinkIn
MakeMIT × Harvard
Smart glasses converting blink patterns into encrypted BLE messages.
Team of 3 · wearable BLE
Typing with your eyes: custom blink patterns map to messages that travel device-to-device as encrypted BLE, read out on an LED four inches from the lens. The team built two working prototypes in one weekend — a full glasses unit plus a second board proving two-way transmission — iterating the enclosure repeatedly to fit sensors and display into a wearable form factor.
Problem
Convert blink patterns into messages in real time.
Build
Built smart glasses that convert blink patterns into encrypted BLE messages with onboard LED status feedback.
Architecture note
On-device blink input maps to encrypted wireless messages with immediate visual status.
CF-07BlockOut
HackDuke
Real-time credit card skimmer detector and automated response system.
Solo build · hardware + AI
A solo build against team fields. An IR sensor catches the skimmer the moment it is attached: the terminal locks, a camera photographs the suspect, Gemini writes a physical description, and an automated AI call reaches police — in seconds, versus hours of manual footage review. The demo skimmer itself was hand-fabricated from failed 3D prints when nothing else fit the terminal.
Problem
Detect and respond to credit card skimmers in real time.
Build
Built an IR sensor and ESP32 system that locks the terminal, captures and identifies the suspect through the Gemini API, and triggers an automated response.
Architecture note
Hardware detection triggers terminal lockdown, suspect identification, and automated response.
Python · FastAPI · React · TypeScript · Swift (iOS) · JavaScript · MongoDB · Full-stack · Unix/Linux
Proof
Shipped FastAPI systems for IRIS and The Underlayer; deployed Counterfeit Eye to iOS.
Data structures and algorithms · Distributed and parallel systems · Networking · Agent/system debugging and observability
Proof
Cut Playback IQ response time from 60 seconds to 10 seconds through parallelized, distributed tool orchestration.
LLM agent orchestration · Multi-model pipelines · LangGraph · Gemini · Real-time computer vision · Fine-tuning with PyTorch and timm · Custom datasets · Structured-output prompt engineering
Proof
Orchestrated three models under 2 seconds in IRIS and fine-tuned Counterfeit Eye on 1,200+ original images.
Reverse engineering · Malware analysis · Cryptography · Network traffic analysis · OSINT · Vulnerability assessment · Internet-scale scanning with Shodan
Proof
Solved all 7 NSA Codebreaker tasks and characterized ~9,900 surveillance recorders.
NSA Codebreaker Challenge 2025USENIX Security '26 accepted poster
ESP32 firmware (C++) · BLE · WiFi · WebSockets · Sensor integration · Rapid prototyping
Proof
Built ESP32 systems for IRIS, The Underlayer, BlockOut, and Counterfeit Mind — BLE/WiFi scanning, NFC, sensors, and displays.
FW-01 · Apr 2026
1,000+ attendees · 3 winning teams
Bitcamp · University of Maryland
Mentored teams at a 1,000+ attendee hackathon on architecture, hardware, and debugging strategy; directly contributed to the technical direction of 3 winning teams.
FW-02 · Apr 2026
Hardware lab · tooling operations
StarkHacks · Purdue
Ran hardware lab and tooling operations at the world's largest hardware hackathon.
Sami's path runs from Morocco to Detroit.
Both places are full of talent and short on infrastructure. He builds anyway.
Mostly self-taught, he learned by building — turning ideas into working hardware and software, usually faster than anyone expects.
Foundation
Largely self-taught
Studying cybersecurity at a college in Detroit while building and shipping real systems.
Languages
Arabic/Darijah [Native] · French [Bilingual] · English [Full Professional] · Spanish [Elementary]
Approach
Prototype fast, ship complete, and make it physically real. Most of these projects went from idea to working demo in a single weekend.