Ali Beikmohammadi
I build AI systems that have to work twice: rigorously enough to publish, reliably enough to run in production.
Stockholm, Sweden
Google Scholar · 14 Sep 2026
About
I am a Senior AI Engineer at Sandvik / Seco Tools and a PhD candidate in Computer and Systems Sciences at Stockholm University, specialising in reinforcement learning, deep learning and federated learning. I work at the point where research and production meet — and I find that point more interesting than either side on its own.
My thesis, Toward Sample-Efficient Reinforcement Learning, asks how agents can learn efficiently when prior knowledge is available, and how federated and distributed methods converge when the convenient assumptions fail — without data similarity, under compressed or clipped gradients, and under model mismatch between clients.
In industry I take that work to production: computer-vision systems for industrial defect detection, LLM pipelines generating content for 35,000+ products, and agentic assistants serving customer, sales, HR and shop-floor users, along with the MLOps practice that keeps them running.
I ranked first in my cohort in both my Bachelor's and Master's studies in Electrical Engineering, at Bu-Ali Sina University and Amirkabir University of Technology respectively. I spent time as a visiting PhD student with the Artificial Intelligence and Machine Learning group at Universitat Pompeu Fabra in Barcelona, and I have collaborated with SCANIA CV AB, Hitachi Energy, SINTEF AS and KTH.
Teaching matters to me: I have taught or assisted on 14 courses and supervised more than 30 Master's theses. What I care about most, though, is the part in between research and delivery — turning an ambiguous problem into something a team can actually build, deploy and maintain. I am always open to research collaboration, teaching, and applied AI work.
Research
Reinforcement learning and distributed optimisation, with an emphasis on efficiency and guarantees.
- Sample-Efficient Reinforcement Learning
- Federated & Distributed Optimization
- Reinforcement Learning for Control & Cyber-Physical Systems
- Agentic & LLM-Based Systems
Engineering
Production AI systems — computer vision, LLM pipelines, agents and the MLOps around them.
- Automated Tool Wear Inspection
- AI-Generated Product Descriptions at Scale
- Domain-Specific AI Agents
- MLOps Platform Standards
Selected publications
-
A Cost-Sensitive Transformer Model for Prognostics Under Highly Imbalanced Industrial Data
-
Human-inspired framework to accelerate reinforcement learning
-
Compressed Federated Reinforcement Learning with a Generative Model
-
Accelerating actor-critic-based algorithms via pseudo-labels derived from prior knowledge
-
On the Convergence of Federated Learning Algorithms without Data Similarity
-
TA-Explore: Teacher-assisted exploration for facilitating fast reinforcement learning
-
SWP-LeafNET: A novel multistage approach for plant leaf identification based on deep CNN
News
- PhD defence at Stockholm University — Toward Sample-Efficient Reinforcement Learning: Theoretical Foundations and Algorithms.
- GICA: The Gap-Index Compositional Arm Framework for Sample-Efficient Test-Time Scaling accepted at TMLR.
- Active Inference for Adaptive Traffic Signal Control in Noisy Nonstationary IoT Environments accepted at IEEE WF-IoT 2026.
- A cost-sensitive transformer model for prognostics under highly imbalanced industrial data published in Cluster Computing.
- Deciphering conductivity in PEDOT guided by machine learning: From solvent baths to charge paths published in Physical Review Materials.
- Collaborative Value Function Estimation Under Model Mismatch accepted at ECML-PKDD 2025 (top 24% of submissions).