Things I've Built

DrakeSort: Autonomous Robotic Trash Sorting via VLM Reasoning

Built a simulation-based robotic sorting system in Drake that automates the identification, grasping, and placement of recyclables from a mixed pile. The perception pipeline uses YOLO11 for object localization and GPT-4o for zero-shot material classification — so the robot can categorize items it's never seen before based on visual reasoning rather than pattern matching.

For manipulation, we combined multi-view point cloud fusion for geometry-aware grasp planning with a dynamic throwing primitive based on analytical ballistic modeling. Instead of slowly placing objects into bins, the robot calculates trajectories and tosses them — decoupling cycle time from workspace size.

Using Knowledge Graph Embeddings to Solve the Reversal Curse

Mitigated the "Reversal Curse" in LLMs by integrating Knowledge Graph embeddings and symmetry-aware training to enable bidirectional logical inference. Developed a model-agnostic methodology that improves reverse-relationship reasoning without requiring architecture-level modifications.

Layer-wise Analysis of FinBERT Embeddings for Enhanced Financial Sentiment

Conducted a comprehensive layer-wise evaluation of FinBERT and BERT embeddings for financial sentiment classification. Used UMAP visualization to inspect intermediate transformer layers and understand how financial domain knowledge is encoded across the network.