Open to ML / NLP roles

Machine learning,
honestly measured.

I'm Rashed — I build ML systems end to end and publish how well they actually work, including when the number is bad.

See the work Get in touch

StockSense — one model, three evaluation conditions

directional accuracy
Random 80/20 split — the number most papers publish 86.87%
Strict temporal split — train before Feb 2026, test after 38.41%
Temporal split, acting only above 0.70 confidence 50.00%

The middle number is the honest one, and it's in the paper. A random split on time-series finance leaks the future into training. Removing that leak cost 48 points — so the system was rebuilt to abstain instead of guess, which is what a trading agent should do anyway.

2peer-reviewed papers published
2papers accepted, in press
1speech dataset paper under review
BScMathematics, before the CS master's
Selected work

Three systems, built end to end.

Each one shipped past the notebook: a dataset, a model, an interface, and an evaluation someone else can check.

Research system + live agent

StockSense

A multimodal framework that reads a financial news headline, looks at the stock's 30-day candlestick chart, and checks 18 technical indicators before committing to a direction. When the three signals disagree, it stays out of the market.

I built the dataset from scratch: 15,578 news articles matched to 6,542 generated chart images and daily prices across 57 tickers, producing 28,598 labelled pairs over a year of trading. It runs today as a Streamlit agent that explains every call.

FinBERTLLaVA / GroqOpenCVXGBoostSQLiteStreamlit
Problem
Text sentiment alone predicts direction barely better than chance — 44.9%.
Approach
Late fusion of text, chart vision, and technical features into one classifier, with a confidence threshold that can decline to trade.
Result
86.87% under random split; under a leak-free temporal split, selective prediction lifts accuracy to 50.00% at 0.667 specificity.
Honest bit
Q1 2026 concept drift is reported in full rather than tuned away.
Applied — clinical research

ICU medication chatbot

A conversational assistant for medication administration, built as the intervention in a nursing master's thesis at JUST studying chatbot effectiveness among Jordanian ICU nurses.

The constraint that shaped it: in an ICU, a wrong or vague answer is worse than no answer. The assistant stays scoped to its reference material and hands off rather than improvising.

Conversational AIArabic / EnglishScoped retrieval
AA Chatbot designed and built by me for Anas Abd Alhafiz, MSc Nursing, JUST.
Client
Anas Abd Alhafiz — MSc Nursing candidate, Jordan University of Science and Technology.
Users
Registered ICU nurses in Jordanian hospitals.
My role
Built the assistant end to end and supported the study deployment.
Status
Used in an active thesis study.
Product

دُعوة — Duaa

A digital invitation platform for the Jordanian and Gulf markets. Guests open a link and get a full-screen invitation — envelope animation, countdown, RSVP — with no app to install and nothing to sign up for.

Every invitation is driven entirely by its URL, so a host can generate and share one without a backend account. Arabic typography renders as real text rather than baked into images, which keeps it sharp and editable.

HTML / CSS / JSURL-driven stateQR check-in
Market
Jordan and the Gulf; weddings and graduations.
Built
Live design studio, template system, RSVP flow, QR check-in at the door.
Hard part
Arabic text over rendered video without losing legibility or falling back to screenshots.
Status
In active development.
Research

Published, accepted, and under review.

Wisdom Journal
Vol. 6(1), 2026
AI-Driven Analysis of Seismic Events: Severity Classification and Epicenter Modelling in the Middle East 99.67% severity accuracy · epicenter R² = 0.96 lat / 0.90 lon
Published
Seismic AI
earlier work
Machine learning for seismic event analysis First paper in the seismic series
Published
Video forensics
Detection of Video Forgery Using Spatio-Temporal R(2+1)D Network Architectures FaceForensics++ (C23) · 86.15% test accuracy · F1 86.14
Accepted
Digital hospitality
Machine Learning for Review Score Forecasting in Digital Hospitality XGBoost · R² = 0.9135 · RMSE 0.2196 across 8 compared models
Accepted
Speech · Arabic NLP
Arabic speech dataset — contributor Details after publication
Under review
About

I came to machine learning from mathematics, and from a classroom.

I studied mathematics first, then spent years teaching it in Jordanian government schools. Teaching is where I learned that a result you can't explain to someone else is a result you don't fully understand yet — which turns out to be the whole job in applied ML.

Now I'm finishing an MSc in Computer Science (AI/ML) at Jordan University of Science and Technology, with a thesis in NLP. The maths background shows up in the parts people skip: how the data was split, what the baseline really is, whether a metric means anything under class imbalance.

Alongside research I run FusionX, where I take on applied AI and automation work — the ICU chatbot above came through it.

I also run a VW ID-series EV diagnostics service: DC fast-charging fault diagnosis, offline firmware flashing with ODIS-E, and live CAN data over OBDeleven — 30+ vehicles so far. It's the same instinct as the research. Read the telemetry, isolate the cause, don't guess.

Now

MSc Computer Science (AI/ML), JUST — NLP thesis

Before

Instructor — Ministry of Education, Udemy; ex-Abwaab

Before

Higher Diploma in Educational Methods, Hashemite University

Start

BSc Mathematics, Hashemite University

Contact

Open to NLP and applied ML roles, research collaboration, and consulting.

The fastest way to reach me is email. Code and papers are on GitHub.