I build LLM agents you can trust — measured with golden evals, per-claim citation checks, and full tracing. Based in Germany. Open to AI / Agent Engineer roles.

🔎 What I'm known for: evaluation & agent-observability rigor — turning "the demo worked" into "here's the eval score, the trace, and the CI gate."

• Lumen — open-source agentic AI tutor. Custom multi-agent orchestrator (no LangChain), course-scoped RAG over pgvector, a 9-tool MCP server (stdio + HTTP) on the official registry, a golden eval suite (LLM-as-judge) wired into CI, and observable per-call traces (tokens, cost, latency). Production deploy on AWS Graviton. ⭐ 66 · live demo · public /eval
• Thoth — open-source agentic systematic literature reviews with per-claim citation verification (cite_check) that scores every cited claim against its source paper and flags hallucinated references. LangGraph agent (planner → assessor → drafter → critic → cite_check), an authenticated MCP server (OAuth 2.1 + PKCE), Langfuse + OpenTelemetry tracing, and a public eval dashboard. 676 tests. live · eval dashboard
• ForgeJudge — open, always-on leaderboard + CI gate for autonomous coding agents. A hand-rolled single-agent solver fixes real bugs from the issue text alone; every patch is scored by execution-as-judge (verified equivalent to the official SWE-bench grading, and cheat-resistant), every run emits a public OpenTelemetry/Langfuse trace, and a multi-seed CI gate blocks any regression — on a $0 stack against a contamination-resistant, mutation-hardened golden set. A model-swap proves the harness is the deliverable (llama-3.1-8b 42% → llama-3.3-70b 83% pass@1, harness fixed). FastMCP server on the official registry. live leaderboard

The numbers are public on purpose. Thoth's dashboard shows citation precision 97% and recall 74% on the goldens scored so far, with faithfulness and coverage tracked in the open as the golden set fills out — a regression is a falsifiable public signal, not a hidden one. Lumen's /eval publishes all three suites whole, strong and weak alike (authoring 3.85/5). The point isn't a perfect score — it's that every score is measured, reproducible, and gated in CI.

AI / Agents: LLM agents · RAG / pgvector · multi-agent orchestration · MCP · LangGraph · Claude API · prompt & context engineering Eval & Observability: golden evals · LLM-as-judge · cite_check · Langfuse · OpenTelemetry · CI regression gates Languages & Backend: Python · TypeScript · Java · FastAPI · Flask · Node.js · PostgreSQL · Redis Infra: Docker · Kubernetes · Helm · AWS · CI/CD

3+ years shipping production software · BSc top of class (Al-Azhar) · MSc Digital Transformation (FH Dortmund) · currently going deep on LLM internals and agent architectures — transformer mechanics, RAG, evaluation, and LLMOps.

Foundations (proof of depth): GPT from scratch (PyTorch) — a 297-test from-scratch GPT (tokenizer → attention → training loop → LoRA). Not just an API caller.

📫 ahmedhobeishy@gmail.com · 🔗 ahmedhobeishy.tech · 💼 in/ahmedhobeishy