Open source & experiments.
Agentic AI applications, fine-tuned models published to the Hugging Face Hub, research codebases, and a PyTorch course watched by the freeCodeCamp audience.
01 — Selected work
Built end to end.
Agentic AI2025
Podcast Summarizer Agent
Agentic AI application that discovers, transcribes, and summarizes podcast episodes on its own.
- Built an end-to-end agentic AI application that autonomously discovers, transcribes, and summarizes podcast episodes, orchestrated as a LangGraph state machine with an LLM reasoning node, a tool node, and conditional routing for retries and error recovery.
- Integrated third-party APIs as LLM-callable tools with dynamic tool selection and structured output handling (SerpAPI search, YouTube Transcript API for automatic speech-to-text captions, MongoDB for persistent storage and metadata management).
- Containerized the service with Docker and deployed it as a REST API with health-check endpoints, a scheduled cron worker, and a Streamlit web interface for non-technical users.
Teaching2024freeCodeCamp
Learn PyTorch in 5 Projects
A full-length applied deep learning course published on the freeCodeCamp.org channel.
- Authored and presented a full-length PyTorch course published on the freeCodeCamp.org channel, teaching applied deep learning through five hands-on, end-to-end projects.
- Covered tabular classification with feed-forward networks, image classification with custom CNNs, transfer learning with pretrained vision models, audio classification, and text classification by fine-tuning BERT.
- Maintained the companion open-source repository of reproducible Jupyter notebooks covering data loading, training loops, and evaluation, used as a public learning resource by the ML community.
LLM Fine-Tuning2024
Text-to-SQL LLM (Qwen2 Fine-Tune)
Fine-tuned Qwen2-Instruct that turns plain-language questions into executable SQL.
- Fine-tuned Qwen2-Instruct on 10K natural-language-to-SQL pairs (DuckDB text2SQL corpus) to translate plain-language questions into executable SQL, enabling natural language interfaces over relational databases.
- Built a reproducible, configurable CLI training pipeline (model ID, dataset, epochs, batch size, learning rate) with checkpointing, step-level logging, and automated run-completion notifications.
- Evaluated with accuracy, precision, recall, and F1, and published the trained model to the Hugging Face Model Hub with a model card, usage examples, and documented limitations for downstream reuse.
TransformersPEFTFine-TuningHugging Face
Research Engineering2026Under review · Q1
Arabic Medical QA: LoRA vs. Full Fine-Tuning
The open-source research codebase behind a Cluster Computing (Springer) submission.
- Open-sourced the research codebase behind the Cluster Computing (Springer) submission: a configurable two-stage pipeline (domain-adaptive continual pre-training, then supervised fine-tuning) spanning a 22-condition grid across 4 open-weight models (Llama-3.1, Qwen2.5, Jais-2, Fanar) with LoRA, QLoRA, full fine-tuning, and zero-shot baselines.
- Implemented a deterministic log-probability evaluation protocol that removes decoding confounds, plus an automated data-cleaning pipeline reporting accuracy, macro-F1, and per-medical-specialty metrics.
- Engineered memory-efficient distributed training (8-bit AdamW, gradient checkpointing, bfloat16) with SLURM job arrays on AWS A10G nodes, and automated experiment tracking, model versioning, and checkpoint publishing through Weights & Biases and the Hugging Face Hub.
02 — From GitHub
More repositories.
Recently pushed public work, synced live from GitHub.
03 — Teaching & writing
Explaining the work.
Courses, open-weight models, and write-ups for people learning applied ML.
Need this kind of system built?
Agentic workflows, RAG over your own corpus, or a fine-tune that actually beats the base model — happy to talk specifics.