kimichen

Kimi Chen

Hi! I'm Kimi, a Data Science and Applied Mathematics double major at UC San Diego with a minor in Cognitive Science. I focus on machine learning, with a particular interest in computer vision, reinforcement learning, and enjoy building things that actually work. (Yes, I'm named after Kimi Räikkönen, the 2007 Formula 1 world champion. Not sure if it says much about me, but it's a great conversation starter.)
Learn more about me here.

Email: kimichen1115 [at] gmail [dot] com

Selected Projects

Disinformation Detection via Various DL Models

For my high school senior capstone, I built and benchmarked deep learning models, including LSTMs and fine-tuned transformers, to detect fake news. I compared their performance against human judgment and found that AI models, especially GPT-4 Turbo, often outperformed people on out-of-distribution news samples. The project revealed both the promise and current limits of AI in combating disinformation.

Read More Code

Saliency from the Sky

As part of Triton UAS, I led the development of a real-time aerial computer vision system for object detection, geo-localization, and large-scale image mapping during fully autonomous drone missions. Our YOLO-based detector, optimized for top-down drone imagery, ran on a Jetson Orin Nano with ONNX acceleration and integrated into a C++ pipeline for GPS-tagged target localization. I also built a robust image stitching module to create high-res maps of the environment mid-flight.

Read More Code

File Order Randomizer for Premiere Pro

Frustrated by Premiere Pro's lack of native file randomization, I built an extension to streamline the creative process for fast-paced editing. Written in ExtendScript and integrated via Adobe CEP, the tool recursively flattens project bins and inserts randomized media clips with precise timing logic. No more manual renaming or external scripts! With over 7000 downloads and a 4.9-star rating, it's still helping editors simplify their creative workflow years later.

Read More Link

Experience

Research Assistant at UC San Diego — Prof. Zhiting Hu's Lab

Mar 2025 - Present
  • Generated 3M+ frame-action pairs with an RL-driven data pipeline that keeps our diffusion-based World Model stable on long roll-outs.
  • Designed an evaluation suite scoring action consistency, controllability, and interaction fidelity for game-centric world models.
  • Enabled vision-language models to play 2-D & 3-D games through the World Model, opening the door to zero-shot agent benchmarking.

Computer Vision Lead at Triton Unmanned Aerial Systems

Sep 2024 - Present
  • Re-architected a 100-file Python/C++ codebase into modular stages, cutting new-model integration time from ~2 days to <3 hours.
  • Fine-tuned and exported a YOLO v11 model to ONNX, achieving real-time detection on a Jetson Nano during live flight tests.
  • Built an OpenCV orthomosaic stitcher for high-res panoramic maps; coordinated with pilots and aero engineers for seamless integration.

Data Science Intern at Lanner Inc.

May - Aug 2023
  • Trained gradient-boosted and time-series models that flagged cost anomalies with 96% precision and suggested optimal reorder dates.
  • Automated data wrangling via Power Automate RPA, cutting processing time by 70%.

Web Developer Intern at Silverline EAS

Apr - Nov 2023
  • Boosted Lighthouse performance & SEO scores by 30 points through caching, image optimisation, and code-splitting.
  • Refactored monolithic CSS into a modular architecture and wired in a CMS for one-click content updates.