Research Experience
Exploring deep learning algorithms for fun since high school! Been interested in systems recently as well. My collaborators and professors are great! :)
Robot Teaching & Teaming Lab | Undergraduate Researcher
- Research under Prof. Mike Hagenow
- Built lab-wide xArm control stack for 11 researchers at 80 Hz with Nix + ROS2
- Trained diffusion policies on hundreds of demos for 16-step action prediction and "third-hand" collaboration
ADAPT @ UIUC | Undergraduate Researcher
- Research under Prof. Charith Mendis
- Expanding TensorRight DSL (verifies ML compiler tensor rewrites) with PyTorch & TensorFlow
- Desugared 46 TASO rewrite rules into XLA-HLO in TensorRight for rank-polymorphic verification
- Added rank preconditions with lower-bound constraints for rule-checking into DSL
- Exploring automated rewrite rule synthesis with TensorRight verification
Stewart Lab @ Morgride Institute for Research | Research Intern
- Supervised under Ron Stewart @ Morgridge Institute for Research
- Published paper in BMC Bioinformatics
- Combined biomedical literature-based discovery with LLM hypothesis evaluation (feel free to play with it here)
- Fine-tuned custom LLM for biomedical text analysis with synthetic data (8,600+ downloads on Huggingface!)
- Improved accuracy of RAG relevance filter from 70% to 90% with NEFTune and RsLoRA
- Deployed custom LLM on university HPC with Nvidia Container Toolkit + Docker & Paramiko
- Sped up internal analysis pipeline by 15x with batching + VLLM
Susceptibility of Adversarial Attacks on Medical Image Segmentation Models | Independent Publication in High School
- Self published at IEEE's 2023 International Symposium on Biomedical Imaging (ISBI) conference. I got cited twice! 🎉
- Explored the efficacy of adversarial attacks on SOTA image segmentation models (UNet, UNet++, UNet + ResNeXt-101 backbone, UNet + EffNet-B7 backbone)
- Discovered that varying the loss used in FGSM improved Attack Success (defined in paper)
Deeply Supervised Transformer Encoders | Independent Research in High School
- Supervised under Prof. Mohammad Taher Pilehvar @ University of Cambridge
- Examined the application of UNet++ style deep supervision in transformer encoders, which unfortunately didn't work very well :(
- Adding Deep Supervision increased BERT accuracy by 5% on the WNLI benchmark