Computer vision and applied machine learning — and the software that carries it.

I build systems that look at the physical world and decide something about it. Right now that means deep learning for industrial inspection at Patil I-Labs. Before that, malware detection research at IIT Roorkee. Throughout, I ship full products, because a model that never leaves the notebook isn't worth much.

Portfolio · LinkedIn · Email

NOW        Computer vision · Patil I-Labs
WORK       Axios · Equiflow · Alertic · DigiGrub
STACK      PyTorch · FastAPI · Next.js · Solidity
ELSEWHERE  ByteXync · CTFs · one unexplained puzzle

Intern–Scientist at Patil I-Labs, building computer vision for Machine Vision Inspection Systems — automated defect detection on railway wheels. Line-scan imagery, geometric correction, wheel boundary extraction, segmentation, and synthetic data for the defect classes reality doesn't hand you enough of.

Four things hold my attention alongside it:

• Multi-view geometry and reconstruction from line-scan and photogrammetric imagery
• Synthetic data for defect classes that barely exist in the wild
• Hybrid static/dynamic analysis for Android malware, continuing the IIT Roorkee work where I co-authored a paper on a hybrid detection framework
• Getting models out from behind notebooks and into APIs someone can actually call

The work I'm proudest of right now — inspection models, malware research — lives in institutional and private repositories. What follows is the public half: things I built end to end, and can show you.

An agent that finds internships, rewrites your resume for each one, and applies on your behalf.

Flask · Next.js · MongoDB · Selenium · Gemini

Job boards don't talk to each other, and a static resume loses to a keyword filter it never sees. Axios scrapes Internshala, LinkedIn and Unstop into one index, scores a resume against a specific posting, rewrites it, and drives the application through a real browser session. A Chrome extension captures postings from anywhere and syncs them back. The hard part wasn't the model — it was keeping three brittle scrapers alive behind one stable API.

Repository →

Equity for open-source contributors, computed from what they actually committed.

Solidity · Hardhat · OpenZeppelin · Next.js

An ERC-20 contract that maps GitHub usernames to wallets, scores contribution — commits, pull requests, reviews, issues resolved, lines moved — and distributes revenue against it. Contributors lapse after an inactivity threshold; only an authorised backend can write scores; the surface is reentrancy-guarded and pausable. Small on purpose: it's the one project here where the rules live in the contract rather than my backend, and can't be quietly changed later.

Repository →

Real-time disaster alerts, community reporting, and an assistant that stays useful under pressure.

Next.js · Flask · PostgreSQL · Prisma · Leaflet · Gemini

Official disaster feeds are authoritative but slow. People on the ground are fast but unverified. Alertic puts both on one map — live regional alerts beside user-submitted reports, with community verification to keep the second kind honest — plus a Gemini-backed assistant for preparedness and response questions.

Live · Repository →

A campus canteen that takes the order before the queue forms.

Next.js · FastAPI · Supabase · TypeScript

Pre-ordering, live menu availability, and stock that decrements as orders land, so the kitchen and the students are looking at the same numbers. A small, sharply-scoped FastAPI service — one route per operation, JWT auth, no cleverness — behind a Next.js storefront and an admin dashboard.

Repository →

The methods I reach for first: CNNs, segmentation, feature matching, homography estimation, multi-view geometry. Then whatever it takes to put the result behind an API someone else can call.

• Vision and learning — PyTorch TensorFlow OpenCV scikit-image scikit-learn NumPy COLMAP
• Backend and data — Python Go C++ FastAPI Flask Node.js PostgreSQL MongoDB
• Interface — TypeScript React Next.js Tailwind
• Malware analysis — Androguard DroidBox Cuckoo Sandbox
• Foundation — Docker Linux Git AWS

I'm secretary of ByteXync, my college's technical club. I wrote a CTF challenge for it once — a pirate-themed web puzzle whose README is deliberately useless as documentation and fairly good as a riddle. It's still up, and I won't be explaining it.

I play CTFs and keep a competitive programming habit, mostly because both punish the same mistake: believing you understood the problem before you read it twice.

Email is fastest, and I answer.

mervin10122004@gmail.com · LinkedIn · Portfolio