Production systems, research infrastructure, competition entries, and open-source implementations. Each card states the problem, the delivered capability, and the result.
Problem. New business requirements used to mean weeks of hand-designing and tuning heuristics.
Result. A modular evolutionary framework in which LLMs propose, evolve, and tune optimization algorithms against executable performance tests. It was deployed to cloud resource allocation and scaling at Huawei, cutting integration cycles from weeks to days.
Problem. Coding agents can change a solver codebase quickly. Knowing whether a change is safe, and whether it helps, is the hard part.
Scope. An on-premises harness that runs agent-made changes to the production MILP solver against correctness checks and performance benchmarks. The program is in its methodology stage; measured results will follow when they can be shared.
Problem. Operations planners depended on consultants for days to run what-if and why-not analyses on production schedules.
Result. A conversational agent over an Advanced Planning System that selects APIs by retrieval and runs scenario analyses in natural language, reducing consultant dependency from days to hours.
Problem. Practitioners describe routing problems in prose. Solvers need a precise variant and constraint set.
Result. An LLM tool that extracts constraints such as capacity, time windows, fleet composition, and pickup-delivery pairs from a description, classifies the variant with per-feature confidence and reasoning, asks clarifying questions across rounds, and recommends a solver.
Problem. The benchmark's largest instances, 1,000 to 10,000 customers, had best known solutions that stood for years.
Result. Our OptVerse-CityU team combined evolved operators with an enhanced AILS-II framework using parallel search, shared elite solutions, and dynamic warm starts. The entry finished first with 51 new best known solutions across 100 instances.
Problem. Keeping up with the literature at the intersection of operations research and AI costs hours a week before any reading starts.
Result. An open-source pipeline that fetches and scores new papers, analyzes each one in multiple passes, detects emerging research fronts, and delivers curated briefings. Adaptable to any research domain.
Process reward model for multi-agent search
Continuous optimization for automated algorithm design
NL4Opt competition data and baselines
Tabu search and genetic algorithm for container loading
Branch-and-price for humanitarian aid distribution
AAAI 2025 benchmark for OR reasoning
Benders decomposition with two subproblems
GA for cutting stock and scheduling (C&OR 2022)
Evolution-guided acceleration cuts for MILP
Research intelligence agent
Learned agents over the local search graph (LION 2026)
Multi-start AILS-II used in the CVRPLIB challenge