I'm an AI & Agentic Systems Engineer at Obelion.AI, where I design autonomous agents that perceive, reason, plan and act on real business problems — and lead the engineering on our code generation and marketplace products.
My work centres on multi-agent orchestration with LangChain and LangGraph: coordinated systems that delegate, retrieve, call tools, and critique their own output before a human ever sees it. I've taken four of these from prototype to production, spanning automated code generation, educational assessment, and financial forecasting.
I came to this from Mechatronics & Robotics, which is less of a detour than it sounds. Closing a control loop around a noisy sensor and closing one around a stochastic model are the same problem wearing different clothes — and I spent a year and a half teaching both to students before building them full time.
Reliability comes from typed state, explicit control flow, and hard failure boundaries — not from a longer system prompt.
Every agent I ship carries evals, groundedness scoring, and trace-level cost accounting from day one.
Control theory taught me to design for feedback, saturation, and drift. Agents need all three.
{"role": "AI & Agentic Systems Engineer","company": "Obelion.AI","focus": ["agents", "RAG", "MLOps"],"stack": ["LangGraph", "LangChain", "FastAPI"],"background": "Mechatronics & Robotics","shipping": true,}
“True agentic intelligence shows up when a system stops waiting to be prompted — when it pursues a goal, notices it was wrong, and corrects course on its own.”
Not a prompt in a for-loop. A typed, observable graph where every node can retry, replan, and justify itself — with the receipts to prove it.
Parse intent, normalise inputs, load session memory
Decompose into a typed task graph with explicit dependencies
Hybrid search over pgvector, reranked and MMR-diversified
Schema-validated tool calls with retries and budget caps
Critic scores groundedness, replans on unsupported claims
Stream cited output, log traces and evals to LangSmith
$ langgraph run supervisor --task "quarterly-risk-report"[00.00] supervisor → decomposed task into 4 subgoals[00.42] retriever → pgvector · 128 candidates → 12 chunks (mmr, k=12)[01.07] analyst → tool_call python_repl → 3 tables parsed[01.88] analyst → tool_call web_search → 6 sources cited[02.41] critic → groundedness 0.94 · hallucination 0.02[02.63] critic → 1 unsupported claim → replanning[03.10] analyst → revision applied, citations attached[03.55] supervisor → done · 5 steps · 18.4k tokens · $0.11▸ report.md written · evals passed 18/18$
Agents fail in ways ordinary services do not. These are the properties I engineer for before anything ships.
Depth where it counts for agentic systems, breadth enough to own a feature end to end — from the control loop to the deploy.
Composable LLM application framework
Stateful multi-agent orchestration
Tracing, evals, and regression suites
Chat, embeddings, structured outputs
Model hosting and fine-tuning
Hybrid retrieval and reranking
LoRA and domain adaptation
Supervisor and critic topologies
Four agentic platforms serving real users — code generation, MLOps, education, and financial forecasting. Built end to end, not prototyped and abandoned.
Agentic Systems · Obelion.AI
An autonomous pipeline that takes a business requirement and returns a tested, deployable application — requirements analysis through to CI.
MLOps Platform · Obelion.AI
A central hub for AI/ML tooling — discovery, resource management, and deployment pipelines for teams building and serving models.
Educational AI · Client engagement
Agentic coursework infrastructure: automated assessment generation, a contextual assistant for staff and students, and automated grading.
Financial Technology · Saudi Financial Group
Forecasting agents for equities and options, wired to subscription billing and a Telegram bot for real-time delivery.
From teaching the fundamentals to shipping the systems — a path through AI engineering, robotics, and the classroom.
Obelion.AI
Lead engineer on the Syntera code generation and marketplace products, building the agent runtime behind both.
Educational Institution
Taught robotics, machine learning, and data science to students and working professionals.
University · In progress
The intersection of mechanical engineering, electronics, and software — and where my instinct for feedback loops came from.
LangChain & LangGraph Specialist
Self-directed
MLOps Engineering
Industry experience
Advanced Python for AI/ML
Professional development
AI Ethics & Responsible AI
Teaching certification
The field moves faster than any curriculum. I read the papers, rebuild the interesting results, and keep a running list of what broke — that list has taught me more than any certificate.
Agentic systems, retrieval pipelines, or an AI feature that needs to survive contact with real users — if it is interesting, I want to hear about it.
Multi-agent architectures, orchestration, and the runtime around them.
Getting models off the notebook and onto a reliable serving path.
Where the model meets hardware, sensors, and a real control loop.
Architecture review, feasibility, and team mentoring on AI systems.
“Innovative solutions require dumb mistakes.”