Things I've built & explored

Research code and pre-registrations accompanying our BlackboxNLP @ EMNLP 2026 submission, Geometric Drift Metrics are Insufficient. Implements CKA, NPS, and isotropy diagnostics; the matched-Frobenius perturbation and matched-corpus MLM controls; the cross-validated linear-probe pipeline; and pre-registration documents for every experiment in the paper.

• python
• pytorch
• faiss
• transformers
• clip
• w&b
• mlflow

github paper draft

A multi-agent advertisement analysis pipeline combining computer vision, vector retrieval, and LLM reasoning to extract marketing and business signals from video. Reaches 70–90% frame deduplication and reduces vision-API cost by roughly 70% while preserving downstream extraction quality. The pipeline behind our ACM MM 2026 paper, AdaFrame.

• python
• claude api
• chromadb
• opencv
• pytorch
• hubert
• ffmpeg

github devpost paper draft

Autonomous mobile robot built during my robotics-research internship at Wabash. Fuses GPS, computer vision, and LiDAR streams for real-time outdoor navigation, reaching 97% obstacle detection and avoidance with ~2 m positional precision in the field. Includes emergency-stop protocols and a configuration-driven control stack.

• python
• ros
• opencv
• lidar
• gps
• pid control

github

End-to-end platform for extracting financial entities from unstructured filings, storing them as vector embeddings alongside a typed knowledge graph, and exposing an intelligent query layer that combines graph traversal with semantic retrieval.

• python
• neo4j
• chromadb
• transformers
• langchain

github

AI-powered system for automatic video advertisement analysis built on top of Google's Gemini 3 multimodal models — extracting structured signals about messaging, pacing, and brand cues from raw advertising footage. A follow-up to AdLovin, swapping in Gemini's native long-video reasoning.

• python
• gemini api
• ffmpeg
• chromadb

github devpost

Sim-racing telemetry copilot. Three agents observe a live sim session — one watches inputs and pace, one watches the car state, one talks to the driver — coordinating to surface coaching cues mid-stint. Built to test whether cooperative agents beat a single oracle on a hard real-time perception task.

• python
• multi-agent
• llm
• realtime

github