Curriculum vitae

University of Utah, Kahlert School of Computing, Salt Lake City, UT, USA

• Advisor: Ana Marasović

The Ohio State University, Columbus, OH, USA, GPA 3.936

Utah NLP (Ana Marasović), University of Utah, Salt Lake City, UT, USA

• Research on mechanistic interpretability of language models, focused on chain-of-thought faithfulness - whether each reasoning step describes a computation the model actually performed, and whether the trace shows the full path to its answer.

NLP Lab (Yu Su and Huan Sun), OSU, Columbus, OH, USA

• Contributed to the Mind2Web-2 benchmark (accepted at NeurIPS 2025, Datasets and Benchmarks Track), a large-scale evaluation suite for long-horizon web agents with human-authored tasks involving search, navigation, and real-world interactions.
• Authored around 15 percent of the benchmark tasks and supported large-model evaluation by running Google Deep Research and OpenAI Deep Research systems.
• Implemented large-scale caching of visited web pages, including manual resolution of CAPTCHA checks, ensuring reproducibility of experiments.

XNAV (student-led project), Columbus, OH, USA

• Co-authored a conference paper on map-matching for GPS-denied navigation, extending an existing algorithm with lane-change detection and benchmarking against IMU- and SLAM-based approaches.
• Designed and prototyped a modular perception pipeline (later refined for AutoDrive), integrating lane-detection models (CLRKDNet) with navigation logic.
• Researched, evaluated, and deployed SLAM models for comparison, ultimately selecting OpenVSLAM for experimental use.
• Contributed multiple sections to the paper and supported collaborative writing and editing.

Buckeye AutoDrive Challenge, Columbus, OH, USA

• The SAE AutoDrive Challenge II is a GM-sponsored international competition where universities transform a Chevrolet Bolt into a Level-4 autonomous vehicle.
• Implemented and deployed a 3D multi-object tracking model (MOT-3D) for autonomous vehicle perception, adapting it to handle static objects and manage detection flicker.
• Re-implemented the ROS 2 perception pipeline from scratch, improving reliability and modularity.
• Optimized inter-node image transfer, reducing latency from ~15ms to ~3ms while eliminating CPU bottlenecks.
• Led migration to a Docker-based perception and control pipeline, coordinating integration with other subsystems.
• Mentored new perception members, developed project roadmaps, and led collaboration within a 50+ member multidisciplinary team.

Imageomics Institute, OSU, Columbus, OH, USA

• Developed open-source tools for large-scale image collection and dataset management, enabling the creation of the BioCLIP-v2 dataset (expanded from 10M to 220M images).
• Designed and implemented a distributed image downloader and supporting preprocessing scripts, handling fault tolerance, parallelization, and efficient storage.
• Tools were released as open-source repositories and adopted by the lab to support large-scale multimodal learning research.

Tech Lead

• Designed and led development of a V2X-inspired safety platform for real-time hazard detection and risk-aware navigation.
• Architected and implemented the backend (AWS/Kafka/PostGIS/TimescaleDB) and guided frontend/UI design, coordinating contributions across multiple teams.
• Developed system for hazard detection and risk profiling of roads and intersections, leveraging spatiotemporal data (S2 cells, historical hazard statistics).
• Defined product vision and technical roadmap, bridging user needs with scalable engineering solutions.

OSU - Intro to C++

• Graded assignments and labs, and held office hours to support student learning.
• Helped students debug code and understand core programming concepts; received positive feedback from attendees.

• Reviewed three submissions focused on large language model memory and multi-turn interaction.