Projects

SmartAPS is an agentic conversational system that transforms how operations planners interact with Advanced Planning Systems. Using tool-augmented LLMs with retrieval-based API selection, it enables natural language what-if and why-not scenario analyses, reducing consultant dependency from days to hours.

Started in 2022, this project aims to democratize operations research by using AI to automatically convert natural language business problems into mathematical optimization models. We established the first benchmark in the field with NL4OPT (presented at NeurIPS 2022 Competition), developed multi-agent systems, created the first reasoning benchmark in OR modeling (ORQA), built agentic systems with high accuracy, and introduced graph-based model evaluation metrics.

Key Achievements: Multi-agent LLM framework achieving 80.8% accuracy in generating mathematical models from problem specifications, human-aligned graph-based evaluation metrics that better align with human judgment, and comprehensive datasets covering 15+ application domains to advance AI-enhanced operations research.

Automating the generation of cutting planes for integer programming by combining Large Language Models with evolutionary search. EvoCut eliminates the need for manual expert design of optimization cuts, automatically generating "acceleration cuts" that enhance solver performance without requiring deep domain expertise.

Key Results: Achieves 17-57% optimality gap reduction within fixed time limits and obtains equivalent solutions up to 4× faster. The framework reliably generates, improves, and empirically verifies cuts that generalize to unseen instances without human intervention.

VRP-Agent is an AI tool that automatically analyzes natural language problem descriptions to identify Vehicle Routing Problem features, classify the VRP variant, and recommend an appropriate solver. Users describe their routing problem in plain English and the system extracts constraints such as capacity limits, time windows, fleet composition, and pickup-delivery pairs.

Key Features: Each detected feature comes with a confidence score and detailed reasoning. An interactive Q&A clarification system asks targeted follow-up questions and refines the analysis across multiple rounds with full conversation memory. Supports identification of a wide range of VRP variants — including CVRP, VRPTW, OVRP, HFVRP, VRPB, and PDP/PDPTW — as examples of the problem types it can recognize.

Designed optimization algorithms for cloud hardware configuration at scale, targeting the assignment and routing of optical connections across distributed infrastructure. The work directly reduced infrastructure costs and improved network performance.

Key Results: Achieved a 15% reduction in optical connections and a 21% reduction in communication latency across distributed cloud infrastructure.

Challenge: Community detection algorithms typically use heuristics with no optimality guarantees. Our research shows that sub-optimal partitions are disproportionately dissimilar to any optimal partition, even when their modularity scores are near-maximum.

Technical Innovation: Bayan solves the NP-hard modularity maximization problem using a specialized Branch-and-Cut scheme with novel triangular constraint-based cuts. For a violated node triple (i,j,k), we partition the solution space via two distinct cut types:

• Left cut (equating membership): x_ij + x_ik + x_jk = 0
• Right cut (disallowing co-membership): x_ij + x_ik + x_jk ≥ 2

This disjunction enables efficient exploration of the feasible space through targeted branching strategies (e.g., supernode replacement for left subproblems).

Empirical Results:

• Accuracy: Consistently ranked top-3 out of 30 algorithms for retrieving planted communities (highest AMI on LFR/ABCD benchmarks)
• Quality: Best median performance across 5 metrics (description length, coverage, performance, average conductance, well-clusteredness) on 1000 networks
• Speed: 3.7× faster than Gurobi IP and 15.5× faster than igraph on average; solved instances unsolvable by alternatives within 4-hour limits
• Critical finding: Standard heuristics achieved global optimality in only 43.9% of 104 networks analyzed

More details coming soon.

More details coming soon.