I teach machines tounderstand thephysical world.
Fifteen years on a single question, asked in progressively different languages: what is this system actually doing, and how would I prove it? The vocabulary changed — conservation equations, then neural networks, now agents reasoning over live industrial data. The discipline did not. These days I lead that work without having stopped doing it.
By the numbers
Years of industry delivery
Production AI, cloud platforms and connected devices across three companies
Citations
Across fuel cells, energy systems and applied machine learning
Peer-reviewed publications
Journals including Energy Conversion and Management and Journal of Power Sources
h-index
Ten papers with at least ten citations each
Doctoral GPA
A perfect score on SFU's 4.33 scale — straight A+ throughout
University course sections taught
Thermodynamics, fluid mechanics, heat transfer and controls
Physics first. Then the model.
The most useful thing I bring to a machine learning problem is rarely the model. It is knowing what produced the data, and therefore which questions are worth asking of it.
I arrived at AI through mechanical engineering — fluid mechanics, heat transfer, energy conversion, CFD written in C++ long before anyone thought to call that work data science. A doctorate turned the same training toward neural networks and hydrogen fuel cells. The order matters: it means I read data as the trace of a physical process rather than as a table of numbers.
What followed was less a change of field than a widening of responsibility. First the models, then the pipelines feeding them, then the cloud architecture carrying them — AWS first, Azure since — and eventually the roadmap and the team. I have been building with large language models since the earliest GPT releases, when using them was something you argued for rather than budgeted for. Leading that work has not meant stepping away from it: I still design the systems, review the code, and answer for the result when it disappoints. Ownership is not conferred with a title. It is what you do on the day the architecture is yours and the outcome is not what you promised.
Technical leadership
I lead AI development without stepping out of it: architecture, code review, sprint planning, and the demo environments sales and marketing take in front of customers. Fifty course sections as a teaching assistant taught me the rest — if the value is not legible to someone outside the discipline, the work is not finished.
Production delivery
Six years shipping: cloud portals built from an empty repository, edge interfaces on Raspberry Pi, agentic AI grounded in live production data rather than a standalone chatbot, and the retry logic and circuit breakers that keep it standing when the platform rate-limits under load. Research that never ships is a hobby.
Research depth
Ranked 1st in my graduating class at the University of Tehran — Iran's top-ranked mechanical engineering department — then a perfect GPA through a Ph.D. at Simon Fraser, fifteen peer-reviewed papers and over 500 citations. I am comfortable at the point where a problem has no known answer and the literature runs out.
Research, then production.
Three chapters, one habit: get close enough to the system that the model stops guessing. Select any role to expand it.
- Architected an agentic copilot that acts on live platform data — assets, sensors, anomalies, alarms, maintenance and risk context — through structured tools rather than free-floating chat. The tool-grounding against the real product is the entire differentiator over a bolt-on chatbot.
- Built the production orchestration underneath it on Azure OpenAI: stateful sessions, chained responses, bounded tool loops, context compaction, and audit/persistence logging — the plumbing that keeps an agent from hallucinating its way into a stuck loop.
- Designed and shipped an automated reporting layer that turns raw vibration, temperature, anomaly and alarm telemetry into a plain-language daily briefing for each asset — what changed, what is at risk, and what needs attention — in place of a dashboard someone has to interpret.
- Extended it into multi-pass reasoning: vector retrieval over prior assessments surfaces relevant precedent, which a second model pass folds into the current analysis, giving a continuity that a single-shot response cannot produce on its own.
- Engineered for Azure OpenAI's real failure modes — rate limiting and HTTP 429s — with retries, circuit breakers and concurrency gating, and defaulted to identity-based access via Entra ID and Managed Identity over embedded API keys wherever the platform allowed it.
- Own AI planning end to end: decomposing initiatives into Azure DevOps work items, running sprint planning, and acting as the primary reviewer for the team's AI-related code and architecture.
- Present AI capabilities directly to Sales and Marketing, translating agent architecture into customer-facing value — the same translation instinct that shows up across every role on this page.
Two disciplines, fully.
Not a machine learning engineer who read a physics book, and not a mechanical engineer who took a Coursera course. Both, to depth.
AI & Machine Learning
From classical estimators on sensor data to agentic LLM systems in production.
Where the model met the machine.
Fifteen papers, and the thinking behind them.
Each entry lists not just the abstract, but what the work demonstrates about how I approach a problem.
Three degrees. Three perfect GPAs.
University of Tehran, then Simon Fraser University. Mechanical engineering into mechatronics into machine learning — each one earned, none of them coasted.
Ph.D.
Applied Sciences — Mechatronic Systems Engineering
Simon Fraser University
Vancouver, Canada
GPA 4.33 / 4.33
Thesis · Modelling and Diagnosis of Solid Oxide Fuel Cells (SOFC)
- A perfect GPA on SFU's 4.33 scale — straight A+ across every graduate course
- Machine learning applied to the optimization and diagnosis of commercial hydrogen fuel cells
- Best Presentation Award, 2nd BC Universities Systems and Control Meeting, University of Victoria (2019)
- Recipient of the SFU Big Data scholarship for real-time analysis of commercial clean hydrogen fuel cell data
- Teaching assistant across ~50 course sections, 2017–2021 — see Teaching
M.Sc.
Mechanical Engineering — Thermal Sciences & Energy Conversion
University of Tehran
Tehran, Iran
GPA 4.00 / 4.00
- Straight A+ across all postgraduate coursework
- Modelling and optimization of energy conversion systems, net-zero buildings and power plant efficiency
- Co-translated 'An Introduction to Heat Transfer' (Bergman, Lavine, Incropera & DeWitt) into Persian
B.Sc.
Mechanical Engineering
University of Tehran
Tehran, Iran
GPA 4.00 / 4.00
- Ranked 1st among ~120 undergraduate students — the top-ranked mechanical engineering department in Iran
- Winner, Faculty of Engineering (F.O.E.) Prize — most distinguished student in the department, three years running
- Straight A+ across all undergraduate coursework
- First exposure to programming: CFD solvers written in C++
~50 course sections
Taught or assisted
Twelve years of teaching, 2009–2021, across the University of Tehran and Simon Fraser University — from undergraduate statics to graduate continuum mechanics. It is the reason I can explain a fuel cell to a hardware engineer, a spectral feature to a salesperson, and a model's failure mode to an executive — without changing the substance.
The problems worth my time are the ones where the physics is hard, the data is messy, and the answer has to survive contact with a real machine.