AI
Foundry
Initiative

Most real-world data lives in relational databases — tables connected by foreign keys — yet LLMs struggle to reason over such structure natively. AiGRASP builds techniques to augment LLMs so they can reliably work with relational databases by transforming them into property graphs. A partnership with the Krenicki Center.

Building on the Spring 2026 causal discovery work, this project deploys LLM agents that autonomously invoke causal discovery tools, interpret their outputs, and engage in iterative dialogue with engineers and plant operators to refine causal models for process optimization and fault diagnosis.

Continuing from the award-winning Spring 2026 project, this semester focuses on scaling the HITL verification platform to additional courses, improving the RAG pipeline, and developing smarter triage mechanisms that route only genuinely uncertain AI responses to human TAs.

A partnership between AI Foundry and the Krenicki Center. This phase advances an agent architecture to support simulator calibration, generating realistic business scenarios that account for complex, real-world operational dependencies.

A Piazza-like platform where an LLM automatically generates answers to student questions from course materials — with Teaching Assistants in the loop to verify, correct, and approve before delivery. Deployed in live AI Foundry coursework and won 1st prize ($1,000) at the Innovate & Connect Showcase.

Design and deploy a production-grade LLM inference API compatible with the OpenAI interface, powered by vLLM or SGLang. Targets low-latency text generation and embedding computation for all AI Foundry projects, integrating open models such as LLaMA and Mistral.

Add LLM-powered agents to an open-source logistics simulator, enabling human-like decision-making in complex supply chain scenarios. Agents interpret natural language commands to optimize operations and predict outcomes. Benchmarked against rule-based systems across cost efficiency and delivery speed.

Build a new open-source logistics simulator from the ground up, then train Neural Algorithmic Reasoning (NAR) models to learn dispatch and management of vehicles, containers, and orders. An LLM assistant enables users to define complex what-if scenarios using natural language.

LLMs possess vast world knowledge — "smoking causes cancer" — that data-driven methods cannot access. This project builds a hybrid framework that injects LLM-derived priors into algorithms like NOTEARS, DAGMA, and DAGPA, improving causal graph accuracy in low-data regimes.

Develop novel graph neural architectures for Mixed-Integer Linear Programming problems, bridging data-driven pattern recognition with provably optimal solution methods. Target domains include logistics, scheduling, and chemical process optimization. Evaluated against Gurobi and CPLEX on large-scale benchmarks.

An open-source LLM agent embedded in Minecraft Java Edition that dynamically scaffolds prompt engineering skills for K-12 students. Using tool calling to manipulate the game environment — generating blocks, teleporting players, spawning entities — the agent enables embodied learning about GenAI concepts.