AI Forge

Purdue's AI Forge is both a principle and an initiative aimed at providing Purdue's undergraduate and master's students with advanced skills in state-of-the-art generative AI techniques, taught by internationally recognized AI experts. The key principle of AI Forge is giving students agency, knowledge, and resources to allow them to forge their own path in AI.

AI Forge Course

This semester-long course provides students with hands-on knowledge on state-of-the-art foundation models for AI.

Students will learn the basics of LLMs; in-context learning and RAG; and multi-modal generative AI models.

This course is open to students with an interest in AI who have Python programming experience and a foundational understanding of basic probability and statistics and linear algebra.

Students can enroll during open registration starting December 8. Contact your advisor for more information.

Course number: CS 39000AIF (crn 41156)
Meets Monday, Wednesday, and Friday at 11:30a-12:20p

AI Forge Projects

AI Forge projects give students the chance to apply artificial intelligence to real-world challenges. From sustainability and healthcare to creative technologies and smart systems, each project is designed to be hands-on, collaborative and impact-driven.

Students work in teams guided by faculty and industry mentors to move ideas from concept to prototype, gaining real experience in research, design and responsible AI development along the way.

More information about projects for the Spring 2026 semester to be announced.

Person sitting in front of having a planning and analyzing with display screen and pointing on the data. (Adobe Stock)

Aligning Risk/Opportunity Preferences with Agent's Decision Making Behavior in the Financial Domain

This project explored whether large language models (LLMs) can reason about real investor behavior and connect it to declared risk preferences. Using a unique dataset that combines survey responses, trading history, and market data from European stock exchanges, students investigated whether AI models can identify investors preferences based on their investment actions. Initial experiments show that predicting an investor’s risk level from their trading behavior is surprisingly difficult, highlighting an opportunity for more advanced AI reasoning. The team proposed and implemented an LLM-based approach for identifying relevant risk-aligned behavioral patterns from raw data and showed that models learning over these patterns can significantly better predict investors’ preferences in real-world market conditions.

Chimpanzee sitting on a branch (Adobe Stock)

SimianSight: Multi-modal Vision-Language Model Framework to Monitor and Predict Ape Behaviors

Wildlife conservation efforts require constant monitoring of habitat conditions, which can be too costly to do at the needed scale. Recent advances in multimodal AI system have the potential to significantly lower that cost through automation. This project explored the first steps in this process - whether AI models can be used to better understand animal behavior in the wild. Unlike past work that relies on manually annotated data for training visual detection systems, this project uses a multi-modal foundation model, analyzing chimpanzee behavior in videos without additional data. Through careful analysis of the model’s capabilities, the students designed a multiple-step reasoning strategy for identifying complex social interactions, emotional cues, and behavioral patterns. Experiments comparing our reasoning strategies to traditional prompting show a more than 20% reduction in model prediction error.

Coworkers having a creative discussion in the office lounge area. (Adobe Stock)

Autonomous Data Exploration Agent

This project explores the hidden social dynamics that shape human conversations, including power, coordination, and influence. While large language models show an impressive ability to interpreting text, these models often struggled to recognize deeper, context-dependent social dynamics. Building socially competent AI systems, that can seamlessly interact with humans require models that can understand and react according to these dynamics. To address this, the project developed an autonomous data exploration agent that identified and reasoned about emerging social concepts across large-scale conversational data. By combining efficient clustering methods with context-aware reasoning, the system uncovered latent power relationships and provided new, interpretable insights into how social influence manifests in conversation.

Factually Grounded and Human-Aligned RAG System for AI Governance and Policy

This project focused on improving how AI systems analyze and summarize complex artificial intelligence policy documents. While many recent retrieval-augmented generation (RAG) systems attempted to align outputs with human preferences, their performance on technical and legal-style policy texts remained limited. To address this, the team developed a pipeline that combined an advanced ColBERT-based retriever with a language model fine-tuned using human preference signals. This approach improved both the accuracy and clarity of generated responses, reduced hallucinations, and produced summaries that remained more faithful to original policy texts across different jurisdictions and legislative contexts.

Photovoltaic modules reflect sunset light and clouds (Adobe Stock)

Paid Voices vs. Public Feeds: How Climate Narratives Diverge Online

This project examined how climate change was discussed and promoted across centralized and decentralized social media platforms. Researchers compared targeted advertisements on Meta (formerly Facebook) with organic, grassroots conversations on BlueSky to understand how each shaped public dialogue around climate change. Posts and advertisements were collected into two separate datasets and analyzed using a framework that combined large language models with the HDBSCAN clustering algorithm to uncover latent themes. These AI-generated themes were then compared with results from traditional topic models such as LDA and BERTopic, revealing platform-dependent patterns and highlighting the influence of targeted advertising on climate change communication.