[ AI ENGINEER ]
Mostafa Arafa
I build AI agents that survive production.
Production-grade agents, RAG pipelines, and automation — engineered for the part most people skip: guardrails, evaluation, observability, and memory.
LangGraph
LangChain
LangSmith
DSPy
OpenAI
Anthropic Claude
n8n
FastAPI
Docker
Pinecone
LightRAG
PostgreSQL
Python
k6
LangGraph
LangChain
LangSmith
DSPy
OpenAI
Anthropic Claude
n8n
FastAPI
Docker
Pinecone
LightRAG
PostgreSQL
Python
k6
What I do
Not just building agents — making them reliable.
Agentic Systems
ReAct agents on LangGraph with stateful memory, multi-layer middleware, tool use, and rate limiting — short-term checkpointing plus long-term procedural memory.
RAG & Graph RAG
Retrieval pipelines with Pinecone, plus knowledge-graph retrieval (LightRAG) for multi-hop relational queries that naive RAG can’t answer.
Production Reliability
The hard part: guardrails, LLM-as-judge, eval suites, token and latency optimization, and full observability with LangSmith. Agents that hold up under load.
Debugging & Experimentation
Reproduce, trace, and fix the gnarly failures others give up on — then push further: prototyping new agent patterns (Graph RAG, DSPy, MCP, middlewares, deep agents) and proving what actually holds up.
Flagship
Production NL-to-SQL ReAct Agent
LangChain
LangGraph
OpenAI
FastAPI
LangSmith
PostgreSQL
A production ReAct agent that turns natural-language business questions into safe, read-only SQL over a 120+ table schema — returning structured answers and optional Plotly charts. Deployed on LangGraph Cloud with Postgres-backed state.
1.00 / 1.00
SQL correctness, faithfulness & tool-usage across a 50-question evaluation suite.
0.0%
Error rate under 200 concurrent users in load testing.
~2,300
Tokens saved per turn through progressive skill disclosure.
Guardrails: code blocklist plus LLM judge, per-user rate limiting, and a self-learning sidecar.
Delivered in a commercial setting; client details omitted. Architecture and results are my own.
[ UNDER THE HOOD ]
How it actually works.
ARCHITECTURE — LangGraph state machine
Router
classify intent
Planner
decompose query
SQL Generator
DSPy-tuned · read-only
Executor
safe execution
Validator
LLM judge · self-repair
↺ failed checks loop back to Planner for self-repair
PERFORMANCE — 50-question eval suite
Correctness
1.00
Faithfulness
1.00
Tool-use accuracy
1.00
Error rate @ 200 users
0.0%
p95 latency
low · stable
tokens / turn
−2,300 saved
before
after
Selected work
Open-source projects I built and shipped.
Retail Analytics Copilot
Local, privacy-first agent answering retail analytics questions by orchestrating RAG + SQL. An 8-node LangGraph state machine with DSPy-optimized SQL and a self-correcting retry loop — 90% benchmark success, fully offline via Ollama.
LangGraph
DSPy
Ollama
SQL
Graph RAG Knowledge Bot
Graph RAG agent over a dense technical rulebook using LightRAG — multi-hop traversal of entity relationships, benchmarked as more accurate than naive RAG on relational queries. Orchestrated with n8n.
LightRAG
Graph RAG
n8n
n8n Automation Portfolio
Eight end-to-end automation solutions: multi-agent orchestration, Human-in-the-Loop approval, RAG with Pinecone, autonomous revision loops, and integrations with Calendly, Google Workspace, and FireCrawl.
n8n
Multi-Agent
RAG
APIs
[ GRADUATION PROJECT ]
Autonomous Formula Student Path Planner
Designed the path-planning algorithm for an autonomous Formula Student race car in Python and ROS — optimal raceline generation to minimize lap time, plus reliable local planning on unknown tracks. Validated in Gazebo / RViz. Graded A+.
Python
ROS
Gazebo
2nd — Formula Student Italy 2023
3rd — Formula Student Italy 2022
[ IN PRACTICE ]
Beyond the benchmarks.
Debugging, shipping, conversing, forecasting, experimenting — the day-to-day a single score can’t capture.
DEBUG & DEPLOY
CONVERSATIONAL AGENT
FORECASTING
R&D — TRYING NEW PATTERNS
Stack
The toolchain I reach for.
Agents & Orchestration
LangGraph · LangChain · DSPy · n8n
Observability & Deploy
LangSmith · LangGraph Cloud · FastAPI · Docker
LLMs
OpenAI (GPT-4 / 5) · Anthropic Claude · Gemini · Ollama / LM Studio
RAG & Vectors
LightRAG · Pinecone · FAISS
Data & Languages
PostgreSQL · MySQL · SQLAlchemy · Python · SQL · C++
Practices
ReAct · RAG · Multi-agent · LLM-as-judge · Progressive disclosure · pytest · k6
About
I'm an AI Engineer focused on building production-ready AI systems that are reliable, observable, and built to solve real business problems. My work spans agentic systems, LLM applications, retrieval, evaluation, and automation, with an emphasis on engineering practices that make AI products scalable and maintainable.
My engineering mindset started long before AI. During my Computer Engineering degree, I developed the path-planning algorithm for an autonomous Formula Student race car, helping the team achieve podium finishes in Italy in 2022 and 2023. Today, I bring that same systems-thinking approach to AI—designing, evaluating, and optimizing intelligent applications from concept to production.
BSc Computer Engineering, Arab Academy for Science & Technology — GPA 3.85.
[ EXPERIENCE ]
The path so far.
AI Engineer
PRESENT
34ml
·
Jan 2026 — Present
Engineer production AI agents and LLM applications from concept to deployment. Specialize in agentic systems, evaluation, retrieval, automation, and performance optimization, with an emphasis on building reliable AI products that deliver measurable business value.
Data Analyst & AI Automation Engineer
Al Faysalia Crew Service
·
Jul 2024 — Sep 2025
Built and deployed AI agents (LangChain, n8n) on internal data to automate customer support — RAG, multi-agent orchestration, human-in-the-loop, and autonomous revision loops across 8 end-to-end solutions.
Path-Planning Driverless Engineer
Arab Academy Motor (AAM)
·
Sep 2021 — May 2024
Led a novel path-planning algorithm in Python for an autonomous Formula Student car — 3rd place FS Italy 2022, 2nd in 2023. Modeled vehicle dynamics and optimized racing lines, validated in ROS / Gazebo.
Head of Media & Executive Secretary
IEEE Alexandria AAST SB
·
Sep 2019 — Sep 2023
Led a 50+ member student branch through roles from Media Officer to Branch Secretary, and ran media & marketing for the 1st IEEE Egyptian Student & Young Professional Congress (ESYP 2022).
BSc Computer Engineering · GPA 3.85
Arab Academy for Science & Technology
·
2019 — 2024
Graduated with an A+ thesis on a general path-planning algorithm for autonomous racing.
Contact
Let’s build something reliable.
Open to full-time AI / ML engineering roles. The fastest way to reach me is email.
© 2026 Mostafa Arafa
AI agents that survive production