Jasmine Shone

Pretraining 300M+ parameter models on 15 TB of physics sim data

Designing and implementing a modified version of the SuGaR algorithm, reducing runtime from 30 m to ~1 m 40 s. Proposing and implementing pointcloud downsampling algorithm which reduces points by up to an 12% average. Created parallelized script to improve training efficiency through data cleaning, detecting 6+% of corrupted data. Implemented baseline for evaluation metric and created easy-to-follow jupyter notebook.

Designed, implemented, and evaluated loss functions based on the I-Con framework. Conducted analysis on experimental results.

Designed and implemented image-pair similarity annotation pipeline. Designed, implemented, and explored different experimental setups and loss functions.

Trained multimodal models utilizing pointcloud encoders, object-wise transformers, and vision foundation models, improving evaluation performance by 27.5%
Created and designed keypoint proposal pipeline with specialized VLM prompting SAM, furthest point sampling, mark-based prompting, RANSAC, and point cloud clustering

Coming up with the initial research question + insight to use multiple additional datasets beyond the ones we were provided in the competition. Also worked on most of the stats/modelling work (Granger Causality, ADF, VAR modelling).

Worked on moving the LLM inference pipeline to a locally hosted version of Llama Vision 3.2 11B and Llama 3.2 90B. Learned about quantization, multi-gpu inference, and prompting techniques.

Developed the p-mRMR algorithm and prompt creation framework, designed and wrote an evaluation suite of test applications and performed manual error evaluation of generated apps.

Worked on designing the project/experiments, training the model, doing evals with resnet, and baselines (GANs)