I build agentic AI systems and the models that keep them honest — typed evidence gates, evals, and audit traces instead of self-report. Senior R&D at Synopsys, putting agents into digital-twin software and rebuilding its hybrid analytics for faster simulation. Off hours I'm building SimPilot, a multi-agent platform that runs real engineering simulations end to end. PhD in computational modeling, with a decade of reinforcement learning, scientific ML, and JAX/GPU training behind it.
Describe an engineering simulation in plain English. SimPilot's tool-using LLM agents plan the work, run it on sandboxed compute, check their outputs with typed validators, and hand back a report you can audit. Long-horizon agent work, automated end to end.
Bringing agents into digital-twin software, rebuilding the hybrid analytics framework from the ground up, and building ML models that accelerate simulation and modeling workflows.
GPU-parallel training and evaluation harness for RL policies and deep sequence models. A week-long PyTorch sweep now finishes overnight. Built for fast iteration, metrics, and reproducible runs.
Fine-tuned T5/BERT QA models and built task-specific evals to track accuracy, failure modes, and data quality. The best run reached >80% accuracy on the target QA task.
Open, runnable notebooks on machine learning, scientific computing, linear algebra, and reduced-order modeling. github.com/mhnaderi ↗
I keep circling the same question: what does it take for an agent to be trusted with work that actually matters? Not a demo — work where a wrong answer is expensive and the ground truth is unforgiving.
That question is why I build the way I do: typed protocols over free-form prompts, evidence over self-report, memory that persists. SimPilot is where I'm pushing it hardest.
Other things I keep coming back to: dynamical systems and chaos, and how to evaluate agents honestly. I also have a soft spot for movies that sit with ambiguity — quiet character studies, strange sci-fi, and anything that makes the ordinary feel slightly unreal.