Aditya Mehta

I spent this summer at Moda, where I:

• Built and launched Layerize, a layer canvas decomposition tool 0→1 (launch, with some cool demos)
• Worked on improvements to Moda's agent harness with ideas like programmatic tool calling and rubric-guided verification
• Worked to automate ^ by building the primitives for an agent “continual improvement” stack: rollout/replay environments, LLM judge/eval, trace intelligence, and metric observability.

At Caltech, I've worked in Georgia Gkioxari's group on evaluating and improving multimodal reasoning for LLMs TWIN CVPR 2026 Highlight Same or Not? Enhancing Visual Perception in Vision-Language Models . Before that, at the Caltech Vision Lab, I combined classical CV with foundation vision models for unsupervised segmentation of microscopy images. I've also enjoyed TAing CS 148a (Deep Learning).

Last summer, I was at Encharge AI working on LLM quantization and building an evaluation stack for quantized checkpoints.