Research questions.
Working ML systems.

Advisor: Ray Friedman

• LLM annotation: Led development of a system that codes negotiation transcripts, reducing estimated annotation cost from $5,000 to $3 per transcript and increasing human–AI agreement from 30% to 80% in our later evaluation. Transcript-level context and repeated-run consistency checks support inspection of uncertain labels.
• Negotiation agents: Built LLM-based simulations with heterogeneous agent behavior to study variation across negotiation strategies and outcomes.
• Research software: Built fine-tuning pipelines and web interfaces for researchers to annotate transcripts, run experiments, and compare models.

The work was presented at the 2025 AI Negotiation Summit at Harvard and MIT and at the International Association for Chinese Management Research annual meeting. Read the project overview.

Designed and delivered computer vision for warehouse stock detection and management, with deployments across approximately 30 warehouses.

• End-to-end perception: Owned camera setup, image collection and labeling, OpenCV/PyTorch model development, and deployment. The system detects storage status and shelf alignment in warehouse operations.
• Efficient inference: Exported models to ONNX and optimized inference for resource-constrained hardware.
• Robotics integration: Connected live stock-status updates to the central warehouse control system coordinating automated guided vehicles (AGVs), with data services and an operator interface. Also improved ROS localization and deployment workflows.

Advisor: Tyler Derr

Studied multimodal learning across vision, graph, and language representations.

• Vision–Language: First-authored Inverse-LLaVA, a 7B model with text-to-vision fusion trained without LLaVA-1.5's separate 558K-example alignment stage. Released model weights, reproducible comparisons across nine benchmarks, and component, supervision, and scaling studies. The instruction-only recipe uses 45.6% fewer task-training examples; benchmark differences are task-dependent.
• Graph–Text: Implemented a Temporal Graph Attention Network (TGAT)–LLM pipeline to generate textual node attributes from interaction histories. Evaluated it on the Venmo social graph; designed a follow-up evaluation for Amazon review data.

Advisor: Markus Eberl

• I improved ancient mortar classification accuracy from 60% to 97% by training Vision Transformers on 10 million images.
• I developed a vision-only provenance algorithm that reduced similarity-analysis runtime from months to minutes. At the time of the original project write-up, a manuscript describing the work was in preparation.

I supported three courses through a Graduate Teaching Fellowship, organizing labs, creating quizzes, grading coursework, and hosting office hours.

Graduate courses:

• DS 5620: Probability and Statistical Inference
• DS 5690: Gen AI in Theory and Practice

Undergraduate course:

• DS 3100: Fundamentals of Data Science

I tutored statistics and graded coursework for Econometrics I in the Master of Finance program.