Nameha
CLIENT
Senior Project, King Abdulaziz University
YEAR
2026
MY ROLE
AI features lead · Brand identity

THE PROBLEM
THE OBJECTIVE
The team
A three-person senior project, supervised by Dr. Abeer Almakky. I designed and led the AI layer, built with Claude: the data pipeline, the two LSTM models, the ensemble, the explainability layer and all three AI agents. My teammates built the backend and the mobile app, designed its screens and helped train the XGBoost and Random Forest models.
The identity
I designed the logo and full brand identity. The mark is an open wallet with two sides: one opaque, one glass. The glass side stands for transparency, so the user sees the whole reason behind every choice the AI makes. The upward arrows stand for growth. “Nameha” means “grow it” in Arabic: the promise of the whole app, carried by the logo’s arrows and the auto-invest feature, Grow for me.








Three AI agents
Manual trading, with AI analysis: before a user confirms any buy or sell, the AI gives a recommendation (buy, sell or hold), a plain-language explanation, four market indicators, the trade's effect on their portfolio and a confidence score. They learn why, not just what.
Auto-invest: the user sets an amount and a duration. The AI builds a personalized plan across suitable sectors and stocks, shows an estimated return range and the sector allocation, and invests it once the user confirms. They can pause or stop it at any time.
Risk monitor: the user sets their own loss limit. The system watches their portfolio and alerts them before they reach it, with suggested options; it never sells on its own.
How it works
Three models (LSTM, XGBoost and Random Forest) each predict a stock's direction.
Their predictions are averaged into one confidence score.
A custom explainability layer ranks four factors behind the prediction: price direction, investor activity, price stability and short-term momentum.
LLaMA 3.3 70B turns it all into a beginner-friendly explanation.
The user reads the reasons and decides. Nothing is executed without their confirmation.
The models were trained on 20 Saudi and 20 US stocks from 2010 to 2024, covering Tadawul, NYSE and NASDAQ.
Key decisions
Explain, don't just predict. A black-box answer doesn't help a beginner learn, so every prediction comes with its reasons.
AI proposes, you decide. Every action waits for the user's confirmation, so they stay in control of their money.
Results
The combined model reached about 62% accuracy in predicting a stock's direction, against a 50% baseline. That's in line with published research on stock prediction: markets are efficient, so any pattern that's easy to spot gets priced away quickly. Individually, XGBoost reached 68.71%, the Saudi LSTM 55.85% and the US LSTM 52.77%.
What I'd do next
Expand from 20 to 100+ stocks per market, so auto-invest can diversify more widely, and replace the custom explainability layer with SHAP once its compatibility issue with TensorFlow is resolved.

THE STACK

Claude

Figma

Groq

LLaMA

Scikit-learn

XGBoost

Python

ChatGPT

TensorFlow

YahooFinance

Finnhub
AI layer built with Claude: an LSTM (TensorFlow/Keras), XGBoost and Random Forest (scikit-learn) ensemble, with LLaMA 3.3 70B via Groq writing the explanations. Market data from Yahoo Finance (yfinance) and Finnhub.