Bassam Saleh

Independent AI & ML Engineer at Independent (Freelance)

Developed a multi-modal deep learning system for violence detection in cartoon videos by integrating visual (CNN-LSTM) and audio (1D CNN + LSTM) pathways. Implemented keyframe extraction, audio segmentation, and fusion-based classification using PyTorch. Built an end-to-end pipeline from preprocessing to model training, evaluation, and Flask-based web deployment for real-time violence prediction. The model classifies video scenes into four categories: bloody, explosions, fight, and non-violence.

Independent AI & ML Engineer at Independent (Freelance)

Developed an end-to-end Retrieval-Augmented Generation (RAG) system for querying PDF documents using OpenAI embeddings and Qdrant vector database. Integrated FastAPI backend with Streamlit UI and automated event-driven workflows via Ingest.

Independent AI & ML Engineer at Independent (Freelance)

Fine-tuned Qwen 2.5-1.5B-instruction model using LLaMA Factory and LoRA for Arabic news understanding, summarization, and multilingual translation. Built a complete fine-tuning pipeline, JSON schemas for structured outputs, and deployed the model using vLLM with performance tracking via Weights & Biases.

Independent AI & ML Engineer at Independent (Freelance)

Developed an AI-powered computer vision system to analyze tennis matches by detecting and tracking players, the ball, and court lines in real-time. Implemented trajectory interpolation, hit detection, and mini-court visualization for detailed player and ball analytics.

Independent AI & ML Engineer at Independent (Freelance)

Developed an autonomous AI research agent for analyzing developer tools using LLMs, LangChain, and Firecrawl. Built end-to-end pipelines for data extraction, reasoning, and automated tech recommendations.

Independent Data Scientist at Independent (Freelance)

Conducted in-depth exploratory data analysis (EDA) of the Olympic dataset to identify participation and performance trends. Performed data cleaning, missing value treatment, and statistical visualization using Pandas, NumPy, Seaborn. Generated insights on gender evolution, medal counts, and athlete characteristics.

Independent AI & ML Engineer at Independent (Freelance)

Implemented a dual-output deep learning model to detect tables and column structure from scanned/digitized documents and converted detected tables into structured Excel files with post-processing for accurate cell boundaries.

Independent AI & ML Engineer at Independent (Freelance)

Developed an end-to-end abstractive text summarization pipeline using Seq2Seq LSTM with attention mechanisms. Generated human-like summaries from moderately noisy datasets and evaluated results using ROUGE metrics.

Independent AI & ML Engineer at Independent (Freelance)

Implemented a full computer vision pipeline for football match analysis: detecting players, referees, and ball; generating player performance metrics, team possession stats, and annotated match footage with overlays.

Independent AI & ML Engineer at Independent (Freelance)

Developed a custom auto-grade engine for the Micrograd project from scratch, supporting basic gradient operations and backpropagation.

Independent AI & ML Engineer at Independent (Freelance)

Explored abstractive text summarization and model evaluation.

Developed a multimodal deep learning system for violence detection in cartoon videos by integrating visual (CNN-LSTM) and audio (1D CNN-LSTM) pathways. Implemented key frame extraction, audio segmentation, and fusion-based classification using PyTorch. Built an end-to-end pipeline—from video/audio preprocessing to model training, evaluation, and Flask-based web deployment for real-time violence prediction. The model classifies video scenes into four categories: violent, explosions, fight, and non-violence.

Developed an end-to-end Retrieval-Augmented Generation (RAG) system for querying PDF documents using OpenAI embeddings and Qdrant vector database. Integrated FastAPI backend with Streamlit UI and automated event-driven workflows via Ingest.

Fine-tuned Qwen 2.5-1.5B-Instruct model using LLaMA Factory and LoRA for Arabic news understanding, summarization, and multilingual translation.

Tennis Analysis System: Developed an AI-powered computer vision system to analyze tennis matches, track players, the ball, and court lines in real time. Implemented trajectory interpolation, hit detection, and mini-court visualization for detailed player and ball analytics.

Micrograd — Custom Autograd Engine in Python: Implemented from scratch a minimal automatic differentiation engine to understand how gradients and computational graphs operate. Supported basic mathematical operations and built a full backpropagation mechanism; extended the core to a simple multi-layer perceptron and validated against PyTorch autograd.

Abstractive Text Summarization: End-to-end pipeline using Seq2Seq with attention mechanisms to generate human-like summaries from noisy datasets; evaluated with ROUGE metrics.

Car Detection System: Combined selective search for region proposals with a fine-tuned VG G-16 CNN to detect cars in images. Achieved ~97% validation accuracy and visualized results with bounding boxes.