Ashvin Jaison

Applied AI · Computer Vision · Neural Decoding

Fourth-year, VIT-AP University

I build AI for problems worth figuring out

I like working on problems I don't know how to solve yet. I learn what I need, experiment until something clicks, and connect ideas from different fields to build systems that actually work.

My work spans computer vision, generative AI, optimization and brain signals, with a focus on systems that can perceive, reason and act rather than simply generate an output.

Photograph to come
Ashvin Jaison

New paper Flat in the wrong coordinates: why Adam-vs-SGD sharpness comparisons mislead

Full report · Page 2

ProjectsPage 2

Illustration

Deep Learning · Research paper · 2026

Flat in the wrong coordinates

Adam looks flatter than SGD, but partly because of how its weights end up scaled, not what it learned. I show why, prove SGD can't do the same, and test it across about 680 networks, up to ResNet-18 on full CIFAR-10.

PyTorch · Optimization theory · Paper written, not yet submitted

Illustration

Computer Vision · 2026

Crowd-density early warning from multi-camera video

Tracks people across cameras with a tracker I built from scratch, measures density against Fruin crowding thresholds, and warns before a space turns dangerous.

Python · YOLOv8 · CSRNet · Kalman filter · SciPy

Illustration

Language Models · Evaluation study · 2026

Does retrieval make medical QA more reliable?

Llama 3.1 8B answers 1,000 PubMedQA questions four ways, from no context to the right paper handed over, with the benchmark’s answer leaks closed off. The question is whether retrieval cuts confident errors, not just raises accuracy.

Llama 3.1 · BM25 · MedCPT · Pipeline built, runs next

Illustration

Neural Decoding · In progress · 2026

Decoding imagined speech from EEG

Telling five silently imagined phrases apart from scalp EEG, on public data, with an evaluation built so leakage can’t inflate the result. A four-person capstone.

EEG · MNE · scikit-learn · Framework built, experiments next

Illustration

IoT · Machine Learning · 2024–25

Disaster prediction on a Raspberry Pi weather station

A weather station for under ₹4,500. A k-nearest-neighbours model classifies the weather, and threshold rules on top warn of flood, drought and cyclone risk.

Python · scikit-learn · Raspberry Pi · ThingSpeak

Notices · Tools of the TradePage 3

Every tool posted here is named in a project on page 2. Drag the board, or scroll it sideways.

The AuthorPage 4

Hi, I'm Ashvin.

I study computer science at VIT-AP University in India and graduate in September 2027. I like building things and learning new ones, even when that means a field I haven't worked in before. I adapt quickly, and so far that's taken me from crowd video and weather sensors to medical question-answering and brain signals.

Lately I'm most drawn to neural decoding: reading imagined speech from EEG. I'm applying to MS programmes in machine learning for Fall 2027 and I'm open to ML engineering roles. If either fits, write to me.

CorrespondenceBack Page

Letters to the author

For admissions committees and recruiters

Email
ashvinj2004@gmail.com
CV
TK
GitHub
github.com/Ashvinj2004
LinkedIn
linkedin.com/in/ashvin-jaison-b283072b4