14 pieces
Numerical methods, open source, the things I build, and the road here.
Linking 10 million noisy business records to 1.7 million reference businesses in 44 hours on one laptop, from 0.940 to 0.988642 across 8,269 teams, and why local validation kept lying to us.
How do you know which differential-equation solver is actually fastest on hard problems? You benchmark them - so I contributed eight index-2/index-3 DAE benchmarks to SciML.
Not every term in a system wants to be integrated the same way. IMEX and multirate methods treat stiff and non-stiff parts differently - here's what I added to SciML's solver suite.
The most valuable PRs I've shipped to OrdinaryDiffEq.jl didn't add features - they deleted duplicated stepping code and replaced it with one generic, tableau-driven path.
From never having written a line of Julia to contributing differential-equation solvers for the SciML organization - here's the path that got me into GSoC.
An offline-first facial recognition and liveness-detection app for Android and iOS - on-device MobileFaceNet inference, blink/smile challenges, and a local SQLite store.
Why a model that runs on a GPU is useless on a factory floor - and how INT8 quantization let me fit real neural nets onto a Raspberry Pi and a phone.
Building an edge-AI predictive-maintenance prototype for Tata Technologies InnoVent 2026 - estimating remaining useful life from sensor data with an INT8 model on a Raspberry Pi.
How I went from building React storefronts to contributing numerical solvers in Julia - and why following my curiosity across stacks turned out to be the best career decision.
A year as a GeeksforGeeks Campus Mantri taught me that the hardest part of community building isn't the events - it's getting the first ten people to show up.
Nobody puts 'fixed flaky CI' on a highlight reel, but learning to read failing pipelines is what turned me from someone who writes code into someone who ships it.
How I started contributing to Meshery - MeshSync, Playwright end-to-end tests, and issue triage - and what a 1,000-contributor cloud-native project taught me about working at scale.
An ethnic-fashion storefront with a product catalogue, admin dashboard, wishlist, live search, and a persistent cart - built with React and Vite, no backend required.
Hitting a 1495 max rating on Codeforces wasn't about memorising algorithms - it was about learning to think under pressure. Here's what stuck.