Muhannad A.

You hire me when you need a Claude agent or LLM pipeline that survives production, not a demo that breaks on the second input. I build the AI layer and the infrastructure around it, end to end in Python, and you work with me directly, not a rotating agency team. Top 10% of Upwork freelancers (Top Rated), 100% Job Success, 9 contracts, every completed review 5.0. Recent builds: - A 7-skill, 18-state Claude Code offboarding engine: 4 sub-agents, hook-enforced guardrails, and a JSONL event log as the single source of truth so every run is auditable. - Commission reconciliation across 7 provider statement formats, each with its own layout, delivered as a Claude Agent Skill and handed over for the client to run on his own machine. - A cross-marketplace product matcher that I audited row by row: 99.4% precision verified on the priority stratum, 96.6% across all strata, with an "abstain rather than mismatch" rule so it stops instead of guessing wrong. - A 21-check Python QA pipeline (20 batch gates plus a cross-batch drift gate) and a two-pass LLM-judge review fleet for a fine-tuning dataset. - A GEO upgrade for a news publisher, built with Claude Code and automated schema verification. - A multi-source water-data pipeline: 3 regulatory databases merged into one audited, queryable product, built solo through an agentic Claude Code workflow. When a no-code tool hits its ceiling (branching logic, state, real error handling), that's where I take over. I rebuild the workflow in code so it does what you actually meant, keeping the parts that worked. I work best on projects where: - The input and output are clear, even if the data in between is messy - There is real API integration, data wrangling, or an AI/LLM workflow involved - You need something built, tested and handed over, not just advised on Why Claude-native matters: it's the model family I design around full time, so I know its failure modes and build around them instead of being surprised by them. If your project needs a different model for one step, I will tell you and wire it in. The goal is a system that holds up, not loyalty to a logo. Based in Cairo (UTC+2, UTC+3 in summer): full overlap with European hours and US mornings. Written updates daily on every contract. Stack: Claude Fable/Opus, GPT Astra/Sol, Python, Node.js, OpenRouter, Tesseract, faster-whisper, Playwright, Pandas, SQL.