OptvEvolve: LLM-driven algorithm evolution in production
Result. Deployed to cloud resource allocation and scaling at Huawei, cutting integration cycles from weeks to days.
Mahdi Mostajabdaveh
AI research & technical leadership
Principal OR/AI Scientist, Kinaxis
Google Scholar: 574 citations (Sep 2026)
I develop AI systems that reason about complex problems and discover better algorithms, and lead their path from research to production.
I am an AI researcher and technical leader with more than a decade of experience in optimization and applied research. My work combines language models, multi-agent systems, and mathematical optimization to improve how systems reason, search, and make decisions. I develop new methods, build the evaluations that test them, and lead their translation into working systems.
At Kinaxis, I lead AI acceleration for a production optimization solver and the evaluation infrastructure for coding agents that work on it. Previously, as Senior Staff Researcher and AI Lead for Huawei's OptVerse solver, I led research on AI for optimization, helped deliver a routing engine serving more than 20 enterprise customers, and developed OptvEvolve for LLM-driven algorithm evolution. Our OptVerse-CityU team won the 2026 CVRPLIB Best Known Solutions Challenge with 51 new best known solutions.
My current research focuses on agent reasoning and adaptation, automated algorithm discovery, and language-driven optimization. I hold a Ph.D. in operations research from Koç University and was a postdoctoral fellow with Prof. Michel Gendreau at Polytechnique Montréal / CIRRELT.
10+ years in optimization and applied research
20+ customers served by production routing
1st place in the 2026 CVRPLIB BKS Challenge
Research themes
News
- Sep 2026SEDIMA accepted as a spotlight at REALM, the Workshop for Research on Agent Language Models at EMNLP 2026.
- Aug 2026Joined Kinaxis as Principal OR/AI Scientist, leading AI acceleration for the production MILP solver.
- May 2026Preprints posted for two papers accepted at LION 20 (Milan, June 15 to 19): COAgents and Latent Heuristic Search.
- Feb 2026Our OptVerse-CityU team won 1st place in the CVRPLIB Best Known Solutions Challenge, setting 51 new best known solutions on instances with 1,000 to 10,000 customers.
- Jul 2025Branch-and-price algorithm for equitable last-mile relief aid distribution published in the European Journal of Operational Research.
Selected Systems
Evaluating coding agents on a production solver (Kinaxis)
Scope. A harness that scores agent-made solver changes on correctness and performance benchmarks.
SmartAPS: tool-augmented LLMs for operations management
Operations planners depended on consultants for days to run what-if and why-not analyses on production schedules.
1st place, 2026 CVRPLIB Best Known Solutions Challenge
The benchmark's largest instances, 1,000 to 10,000 customers, had best known solutions that stood for years.
Selected Publications
MASPRM: Multi-Agent System Process Reward Model
arXiv preprint arXiv:2510.24803 (2025)
A process reward model that scores intermediate multi-agent messages and guides step-level beam search and MCTS.
EvoCut: Strengthening Integer Programs via Evolution-Guided Language Models
arXiv preprint arXiv:2508.11850 (2025)
An evolutionary system that discovers reusable acceleration cuts from symbolic MILP models and evaluates them through solver experiments.
Evaluating LLM Reasoning in the Operations Research Domain with ORQA
Proceedings of the AAAI Conference on Artificial Intelligence (2025)
A benchmark for testing whether language models can apply expert operations-research knowledge and multistep modeling reasoning.
Optimization Modeling and Verification from Problem Specifications Using a Multi-Agent Multi-Stage LLM Framework
INFOR: Information Systems and Operational Research (2024)
A multi-agent pipeline that builds optimization models from problem specifications and checks the resulting formulation in stages.
NL4Opt Competition: Formulating Optimization Problems Based on Their Natural Language Descriptions
NeurIPS 2022 Competitions Track, PMLR (2023)
The benchmark, dataset, and shared tasks for translating natural-language optimization problems into executable representations.
SEDIMA: memory for search agents
REALM @ EMNLP 2026 · Spotlight
A memory system that carries useful search experience across runs.
Latest Post
Building a Research Intelligence System
How I automated arXiv monitoring into a four-layer AI pipeline that reads papers, detects emerging research fronts, and delivers daily briefings. Adaptable to any research domain.
Contact
Reach out for collaborations, research opportunities, or just to connect.