Additional Projects

More explorations, learning projects, and tools I've built along the way

Issue Resolver Agent

Autonomous agent that reads GitHub issues, traces root causes, generates fixes, runs tests, and opens PRs — no human in the loop unless escalation rules trigger. Agent identity in SOUL.md, hard constraints in RULES.md: all behavior is version-controlled and reviewable as diffs. Built for the Lyzr AI GitAgent challenge.

Python • Anthropic API • GitHub CLI • Agent Behavior as Code

moms-verdict

Bilingual (English + Arabic) product review synthesis for the Mumzworld e-commerce platform. Two independent LLM calls per language. Every claim cites specific review IDs — invalid citations stripped post-generation. Deterministic refusal logic in code prevents hallucination when evidence is sparse or in a regulated category.

Python • OpenAI API • Anthropic API • Citation Grounding • Bilingual NLP

Transformers for Computer Vision

173-commit deep dive reimplementing vision transformer architectures from scratch in PyTorch — ViT, DETR, Swin Transformer, SAM, TimeSformer, Flamingo, DDPM, GANs, VAE/VQ-VAE, JEPA, and DeiT. Built to understand the internals, not just use the APIs.

PyTorch • ViT • DETR • Swin • SAM • DDPM • GANs • VAE • From Scratch

RAGCraft

Hub for RAG patterns: advanced chunking strategies, hybrid search (dense + sparse), re-ranking, Corrective RAG, Self-RAG, and multi-vector retrieval. Covers Weaviate, Pinecone, FAISS, and Chroma with hands-on implementations.

LangChain • LlamaIndex • Weaviate • FAISS • Hybrid Search • Re-ranking

Multiple PDF Chat App

Conversational AI application for querying multiple PDF documents simultaneously using RAG architecture. Features semantic search, context-aware responses, and document source attribution.

LangChain • RAG • Streamlit • OpenAI • PDF Processing

LLMOps

Comprehensive collection of LLMOps concepts and implementations covering prompt management, evaluation pipelines, model versioning, and production deployment patterns.

LLMOps • Prompt Engineering • Evaluation • Deployment

Reasoning LLMs from Scratch

Implementation of reasoning LLM in PyTorch from scratch, step by step. Explores chain-of-thought, tree-of-thought, and other reasoning mechanisms in language models.

PyTorch • Reasoning • LLM • From Scratch

ML/DL/NLP Learning

Comprehensive collection of machine learning, deep learning, and NLP implementations covering fundamental algorithms to advanced architectures.

ML • DL • NLP • PyTorch • TensorFlow

TorchQuest

Deep dive into PyTorch from fundamentals to advanced techniques. Guide for learning, experimenting, and mastering deep learning with PyTorch from scratch.

PyTorch • Deep Learning • Neural Networks • From Scratch

InsureML Pipeline

End-to-end MLOps vehicle insurance prediction system achieving 87% accuracy. Features MongoDB Atlas, AWS (S3, ECR, EC2), Docker containerization, FastAPI, and CI/CD with GitHub Actions.

MLOps • FastAPI • MongoDB • AWS • Docker • CI/CD

MLOps Learning

Comprehensive MLOps concepts with 130+ commits covering Docker, Kubernetes, DVC data versioning, MLflow experiment tracking, CI/CD pipelines, and Prometheus & Grafana monitoring.

Docker • Kubernetes • DVC • MLflow • CI/CD • Prometheus

AgentForge

Comprehensive guide for building AI agents using modern frameworks. Covers CrewAI, LangGraph, AG2 (AutoGen), LlamaIndex, smolagents, and more with hands-on examples.

CrewAI • LangGraph • AG2 • LlamaIndex • AI Agents

Kubrick AI Course

Working through the Kubrick AI course: building MCP multimodal agents for video processing. Learning Pixeltable, FastMCP, Groq-powered agents, and Opik observability integration.

MCP • Pixeltable • FastMCP • Groq • Video AI • Opik

TinderMuse

AI-powered dating profile image generator using Google's Gemini 2.5 for intelligent image enhancement, background replacement, and style transfer. Streamlit app with privacy-first design.

Gemini 2.5 • Streamlit • Image Generation • Style Transfer

ML Meme