Charvi Bayana

Charvi Bayana

Not another 'CS student'


About

I'm a B.Tech Computer Science student at Mahindra University, Hyderabad (class of 2028), focused on machine learning and computer vision. I head the AI/ML Committee at Enigma, our university's technical society, and lead the ML/CV division of our AUV team, where we're building an autonomous underwater vehicle for SAUVC 2027. Here at the DevFest to learn from the community, meet people building in AI, and share what we're working on. Always happy to talk about computer vision, autonomous systems, or running student tech communities.

Skills
Python
Experience

Software Development Intern

JSW infrastructure

Jun 2026 to Jul 2026
  • Owned end-to-end delivery of a client-facing greenhouse gas reporting platform (Flask, SQLAlchemy, PostgreSQL), from PRDs and BRDs to weekly client calls
  • Built workbook-based data entry, site-scoped access control, reviewer approval workflows and audit logging

Product Engineering Intern

UrbanDart

Dec 2025 to Jan 2026
  • Engineered a distributed valet operations platform with QR pickup verification, OCR plate recognition, a 5-state workflow engine and real-time analytics
  • Built a configurable form platform with conditional branching, submission analytics and CSV export
Projects

Self-Supervised Image Super-Resolution

Trained a custom SRResNet (PixelShuffle, residual learning, perceptual loss), fully self-supervised via a randomized degradation pipeline. An ablation on degradation realism improved PSNR from 24.97 to 25.61 (SSIM 0.647 to 0.675).

Python
PyTorch
CNNs
VGG16 Perceptual Loss

TerraCrop, On-Orbit Satellite Crop Intelligence

Built the autoencoder compression pipeline for a 3-person satellite crop-segmentation system, cutting downlink size 466,000x (about 700 MB to 1.5 KB). Also built the Streamlit demo and CLI integrating the team's fine-tuned TerraMind model.

Python
PyTorch
Convolutional Autoencoder
Streamlit

EEG-to-Image Decoding (NeurIPS 2026 EEG/EMG Foundation Challenge, Track 1)

Fine-tuned the REVE EEG foundation model, aligned to frozen DINOv2 image embeddings with a CLIP-style loss, to retrieve the viewed image from a single EEG epoch. In multi-seed ablations, REVE roughly doubled EEGNet's top-5 accuracy (17.4% vs 8.4%, 3 seeds).

Python
PyTorch
REVE
DINOv2
EEGNet
NeuralBench
Education

Computer science

Mahindra University

Aug 2024 to May 2028