Publication
CrossView: Can Vision-Language Models Reason Across Cameras?
European Conference on Computer Vision (ECCV) 2026 · Accepted as a poster
Sahil Shah, S P Sharan, Harsh Goel, Manvik Pasula, Adithya Hebbalae, Minkyu Choi, Sandeep P. Chinchali
Loading

I like building things and solving problems.
I started coding at 8, making a game in Scratch. The way I learn has not changed much since. Pick something I do not understand, build with it, and find out where it breaks.
I am at UT Austin now for electrical and computer engineering, mostly computer architecture and embedded systems, and most of my week goes to a research lab.
The projects below are the rest of it. A few are useful to other people. Most were the fastest way for me to understand something.

Problems I have been stuck on
UT Austin ECE, Chinchali Group
Robotics Lab Researcher · Sept 2025 – Present
The lab needed a way to tell whether a vision-language model actually reasons across cameras or just answers from whatever single view it was given. Building that test meant matching the same person between cameras in multi-gigabyte surveillance video, then turning those matches into questions with known answers.
The first version handed generation to an LLM, which was the expensive way to do it. Pushing the work down into a symbolic layer instead cut the cost by about 96%, and it took ten versions of the pipeline to settle. I wrote the question set for the MEVA subdataset and presented what we found to the AFRL sponsor.
University of Utah
Student Researcher · Jun – Sept 2024
Undersampled MRI scans finish faster but come back with artifacts. The usual correction trains a network on ground-truth reconstructions. Deep Image Prior does not need any, because the structure of the network does the regularizing, so it runs unsupervised on the scan in front of it.
My work was on convergence: reaching a good reconstruction in 40% less training time, checked with PSNR and SSIM.
Publication
European Conference on Computer Vision (ECCV) 2026 · Accepted as a poster
Sahil Shah, S P Sharan, Harsh Goel, Manvik Pasula, Adithya Hebbalae, Minkyu Choi, Sandeep P. Chinchali
Building, learning, exploring
Degree planning is a constraint problem, so the solver is code and the LLM only explains.
A zero-dependency Node pipeline that commits its own raw data to git as the audit trail.
An MV3 extension and a headless Pusher-WebSocket worker, both firing on the same event.
Where I learned that the shape of the input decides which model you can use.
A five-state game loop around a single audio element.
Small enough that the only real design decision was the balance check.
The first models I trained where I could see why one beat another.
A joke with a build step. The No button runs away; the Yes button ends it.
Undo designed in before anything got fast. Duplicates found by hash, since names lie.
I built the review layer that killed my own headline result. Then it happened again.
Nineteen questions that never name a club, scored by cosine similarity.
Slavomir Durej and Damon Downing’s widget. 13 of its 252 commits are mine.
Static by default, React only where a page needed state, one Worker for the contact form.
Coursework with a fixed spec. The boards still came back from a fab house and had to work.