AI agent is built to manage Gmail, Google Calendar, and Google Meet, automating communication and
scheduling tasks. It summarizes email threads, sends context-aware replies, and schedules meetings.
Integrated with Telegram for real-time conversational control, enabling full automation of routine admin
tasks and boosting productivity.

https://www.twine.net/signin
Presented a beat-based autoencoder model that maps PPG signals to single-lead ECG for efficient on-device
use. Employed a two-stage clustering method for data cleaning and beat segmentation to minimize onset
detection errors. Implemented a subject-independent training protocol to ensure robust generalization.
Achieved high reconstruction accuracy on cleaned MIMIC II data, with a correlation coefficient of ≈ 0.92
and MSE of ≈ 0.0086.

Developed an AI-powered pipeline to extract and process information from scanned documents. Leveraged
Optical Character Recognition (OCR) to convert scanned images into machine-readable text, then applied a
Large Language Model (LLM) to analyze the extracted content and identify key information fields. The
structured data was stored in a database for downstream automation and reporting. This solution enabled
efficient handling of unstructured document inputs, significantly improving data accessibility and reducing
manual processing time.

Built a Retrieval-Augmented Generation (RAG) system to answer user queries about Netflix’s 2024 financial
report. Implemented contextual retrieval for precise document chunking and response generation. Used
Supabase as the vector database to store and retrieve embedded financial data. Enabled accurate, explainable
financial insights through natural language queries to support informed decision-making.

Developping a cyberbullying detection model using a Kaggle tweet dataset. Performed extensive data
cleaning and fine-tuned RoBERTa-large for accurate classification, achieving 1% accuracy improvement on the
evaluation set.

Fine-tuning a BERT-based model to classify Arabic customer reviews from Talabate. The system identified
sentiment and categorized negative feedback into actionable types to aid business decisions. Trained on a
small dataset (2,500 samples) and achieved 74% accuracy on the test set.