I build agent systems in Python, and the mechanisms that check them.
United Arab Emirates · Claude Agent SDK · MCP · Claude Code · FastAPI · Postgres
| reef | Token-frugal agentic coding pipeline for Claude Code. Retry caps that advance as a side effect of the action, author ≠ verifier enforced by a content hash, and an enforcement map that states which rules are code and which are still prose. v0.3.0 · pilot, n=1 · MIT |
| review-toolkit | Three Claude Code subagents reviewing plans, code, and security against a project's own documents, in a fresh context. Ships with a planted-flaw suite so "they catch what they claim" is checkable, including where they don't. 11/11 plants on one model · 3 rules ship untested · MIT |
| wakeel | Arabic-first AI receptionist for GCC small businesses. FastAPI, Postgres, Qdrant, Redis, arq workers, Claude. pre-alpha · MIT |
- A rule that nothing enforces is prose. Every gate I care about gets a test that shows it going red on the defect it exists to catch.
- Author ≠ verifier. Reviews run in a fresh context, because a finding doesn't transfer to the finder that wrote it.
- A green check is a claim too. What ran, on what, and what still ships untested goes in a log next to the badge.
Not refreshed yet — run the profile workflow once.
Earlier work — RAG, ML, and embedded repos from before the agent work
- RAG-QA-System-Application — RAG Q&A over documents: FastAPI, LangChain, Qdrant, RAGAS evaluation. (unfinished)
- mini-RAG — small RAG over my own documents and books.
- customer_churn_chatbot — LangChain agent over churn data with XGBoost predictions and a Streamlit UI.
- Smart_library_system_w_AI — natural-language queries over a MySQL catalogue from a CLI.
- Deep-Learning-with-PyTorch · ML_Projects — course and study-group work.
- Embedded-System-Online-Diploma — C and embedded systems, 2021.
Happy to talk about agent architecture, Claude Code workflows, and how you verify any of it.