MobileNetV2 transfer learning reached 91.97% accuracy and 91.76% macro F1 on 137 held-out images. My from-scratch CNN collapsed to a single class at 20.44%, which is chance for five classes.
Computer Engineering · American University of Kuwait
Hamza Hussein
AI/ML, full-stack web development, computer vision & embedded systems.
I build things that have to be measured, not just demonstrated: a MobileNetV2 classifier checked against a custom CNN on 137 held-out images, and an ESP32 parking rig that will not raise a gate on the word of a single sensor.
I
About Me
Senior Computer Engineering student at the American University of Kuwait, on track to graduate in January 2027 with a 3.1/4.0 GPA. I work across AI/ML, computer vision, full-stack web development and embedded systems, and most of what I have built sits at the seam between models and hardware: a model that has to be evaluated honestly, and a board that has to keep reading its sensors correctly.
On the software side that has meant TensorFlow and scikit-learn: a custom CNN benchmarked against MobileNetV2 transfer learning across 912 audited vehicle images, and a linear regression price model on used-car data that reached a test R2 of 0.699. On the hardware side it has meant ESP32 and Arduino: ultrasonic and IR sensing, automated gate control, an on-site LCD readout, and a local Wi-Fi web interface serving live parking availability with no cloud service behind it.
More recently the work has moved toward generative models. I am using parameter-efficient fine-tuning, LoRA, QLoRA and PEFT, to adapt large pretrained models on a project I am not able to detail, and I hold an IBM SkillsBuild credential in agentic AI. It is the same idea as the MobileNetV2 result above, at a scale where retraining everything is not an option.
The brass numerals above are field notes. Press one.
II
Technical Skills
The inventory below is the same set that populates the astrolabe at the top of the page. Turn a ring up there to read what each one was actually used for.
01
Programming
- Python
- C / C++
- Java
02
AI & Data
- TensorFlow
- Computer Vision
- CNNs
- Transfer Learning
- XGBoost
- LSTM
- LoRA
- QLoRA
- PEFT
- Generative AI
- Agentic AI
- pandas
- NumPy
- scikit-learn
- Regression
- EDA
03
Embedded & Automation
- ESP32
- Arduino
- Ultrasonic Sensors
- IR Sensors
- Wi-Fi
- Bluetooth
- Data Acquisition
04
Tools & Security
- Git
- MATLAB
- VS Code
- Google Colab
- Linux
- Web Security Fundamentals
05
Full-Stack Web & APIs
- JavaScript
- React
- HTML
- CSS
- Node.js
- Express
- REST APIs
- MongoDB
- Stripe
- Vercel
- Render
- Serverless Functions
III
Experience
-
Student Life Office Assistant
American University of Kuwait Part-time
- Supported daily office operations and assisted students with administrative requests in a fast-paced environment.
- Coordinated student events, logistics and communications across teams while maintaining accurate records.
to
IV
Certifications
-
Web Development with AI
CODED Academy, UniCODE Program
A two-week course in the Web Development with AI track of the UniCODE Program, completed with a submitted final project. Verify the certificate
-
Make Agentic AI Work for You
IBM SkillsBuild
Agentic AI: systems that plan, call tools and act toward a goal rather than returning a single response. Verify the badge
-
Inside a Cyber Attack
Sentra Academy, in collaboration with the AUK Cybersecurity Club
A hands-on cybersecurity workshop covering web security fundamentals, offensive techniques, vulnerability discovery and practical attack scenarios through guided labs. ID SNT-8HBZ-S832
V
Selected Projects
An ESP32 reading ultrasonic distance and IR presence together before the gate moves, with live availability served from a local Wi-Fi web interface and mirrored on an LCD.
Linear regression explained 69.9% of the variance in log price. Converted back to money that is a typical error of 32%, and the residual plot is what showed me the error had structure rather than being noise.
Four years of readings from one house. My first run scored R2 0.999 because the features contained the answer; with the same-hour meter readings removed it is 0.598, and XGBoost beats the LSTM.
A full-stack, Arabic-first store with Stripe checkout. Type a sentence like "big screen, under 300 KD, cheapest first" and an LLM turns it into search filters, and prices carry three decimal places because the Kuwaiti dinar divides by a thousand, not a hundred.
Visit the live site (opens in a new tab)VI
Ask About Me
An answering desk wired to my actual CV. It asks a language model on OpenRouter, grounded in the same real projects and results written up on this page and told to refuse anything they do not cover. If the model is unreachable or busy, the desk falls back to a fixed set of answers matched locally in your browser, so it always answers something.
Enquiry desk
VII
The Sky Over Kuwait
The astrolabe at the top of this page is an instrument for reading the sky. This one reads the real thing. Live conditions over Kuwait City come from OpenWeatherMap, and that reading is handed straight to a language model on OpenRouter, which returns a single line about it. Neither key is in this page: the browser calls two endpoints on this site, and those hold the keys in environment variables.
Reading the sky...
Written from that reading
VIII
Contact
Open to internships and graduate roles in AI/ML, computer vision, full-stack web development and embedded systems, in Kuwait or remote. Graduating January 2027.
- hamza-hussein-076a17246
- Location
- Salmiya, Kuwait