Abdelrahman Youssef. Applied AI Engineer. I evaluate frontier models and build agentic systems. Currently contracting at Outlier and Alignerr. Previously led the engineering team behind a production platform on AWS serving 12,000 users.

An agent harness that runs AI coding workers against a task backlog. Each iteration spawns a worker in a fresh git worktree, verifies the output, and merges back. State is persisted before every action — a crash at any point resumes cleanly. Roadblocks escalate via Discord.

Three-route agent orchestrated via a LangGraph state machine with dynamic re-ranking: FAISS vector store (100K+ parent-child chunks, BGE-m3 embeddings, 13 business datasets), Python data analysis with self-correcting code generation (3-retry LLM critique loop), and Tavily web search fallback. Llama 3.2 Vision for multi-modal document understanding; built-in hallucination detection, document grading, and answer validation gates. Evaluated retrieval quality via RAGAS (faithfulness, answer relevancy, context precision, context recall). Containerized with Docker + FastAPI, deployed on AWS EC2 with automated health checks via GitHub Actions CI/CD. Built for consulting-firm knowledge-assistant use case; production-graded: containerized, CI/CD-deployed, instrumented with health checks.

Automated benchmarking framework evaluating 10+ LLMs (GPT-4o, Claude 3.5, Gemini Ultra, Llama 3 family, Mistral variants) against Stockfish 16 across 25 tactical chess metrics: mate-in-1 accuracy, forcing moves, blunder rates, style consistency, stagnation detection, tactic prioritization. Dual-signal evaluation — outcome (win-rate) + reasoning quality (move quality vs. engine) — enabling cross-model-family capability gap analysis. Chess as reasoning proxy: deterministic ground truth (Stockfish), quantifiable sub-goals (mate-in-1, forcing lines), measurable style — separates outcome quality from reasoning quality, a distinction most benchmarks conflate. Open-sourced with a reproducible analysis notebook; demonstrates evaluation infrastructure design without relying on existing benchmarks.

1,000 simulated cars trained in parallel using a feed-forward neural network (tanh neurons, implemented from first principles in JavaScript for browser-side real-time simulation) and a genetic algorithm with mutation-based evolution. Collision detection via segment-intersection math; emergent lane-following and collision-avoidance behaviors discovered without any hard-coded rules. Demonstrates core DL fluency (forward pass, backpropagation, gradient descent, mutation-selection loops) independent of framework convenience — the same primitives underpinning production PyTorch/TF workflows. Live browser demo.