Dev Varshney

Chip Design Intern CaretEDA

I designed an eight-channel EEG neural co-processor and carried it from register transfer level through to a signed-off GDSII layout on SkyWater's open 130nm process, running synthesis, floorplanning, placement, routing, and timing closure through OpenLane. The final layout came out signoff clean at 2.74 mm² across 87,763 standard cells, with timing closed at 60 MHz and both DRC and LVS passing. I wrote the hardware specification and the RTL myself.

Getting there took weeks of iteration, but the last full pass from specification to signed-off layout ran in a few hours, using CaretEDA's agentic AI tooling alongside the Microsoft Discovery platform. It was my first time touching electronic design automation at all. I presented the work at the Design Automation Conference in 2026 at the Microsoft booth, and was given the Microsoft Discovery Young Innovator Award for it.

Product Analyst Intern Scrollmark

I built the quality assurance process for SocialGPT before launch. Over the summer I found more than twenty-five bugs across three platforms and kept an issue tracker that mapped where failures clustered by severity. I reported findings weekly to the marketing team, and worked with the backend engineers to get six of the critical ones fixed.

SSVEP speller for assistive communication Independent

A user looks at one of several squares flickering at different rates, and the system works out which one from activity over the visual cortex. I record on an eight-channel EEG board at 125 Hz, run preprocessing and artifact rejection in MNE-Python, and train a classifier to map the frequency response back to the intended target. The point of building it is communication for people with locked-in syndrome or late-stage ALS. That is a long way off from where this is, but the decoding problem itself is tractable at this scale, and getting the hardware cost down matters as much as the accuracy does.

Multiple sclerosis gene expression analysis miRcore

I worked through a multiple sclerosis expression dataset in GEO2R, then used STRING and KEGG for interaction and pathway annotation. The hypothesis I ended up with connects downregulation of PSMD5 to proteasome dysfunction, and from there to impaired MHC class II antigen presentation. I presented it to around seventy researchers and faculty and I'm continuing into the year-long program.

Bibliometric comparison of invasive and noninvasive BCI ThinkNeuro

Using R against a corpus pulled from Web of Science and Scopus, I looked at how ethical and biological concerns are distributed across invasive versus noninvasive recording literature. I generated the figures and co-authored the poster and abstract, which we presented at the closing symposium.

Molecular Imaging Mini-Fellowship Stanford School of Medicine

I wrote a review on molecular imaging for early detection of neurodegenerative disease, focused on which methods have any realistic path to being cheap enough for routine screening rather than which ones perform best in a well-funded center. Faculty recognized the report at the close of the program.

Capstone on mental privacy and inner speech decoding International Youth Neuroscience Association

Selected for the capstone research track. I took a 2025 study on decoding inner speech and read it against how the same result was covered in the press, which turned into an argument about what mental privacy protections would actually need to look like if this technology keeps working.