// system online
Fullstack Engineer Founder CV Research
CSE at IIT Bombay. I've taken a subscription platform to 450M users at Jio, been founding engineer at Flek AI, and built Palet — a website builder with an AI that writes real code. Somewhere in between, a MICCAI paper on stain normalization and a patent on content-based media timestamps.
press ⌘K for command palette · ` for terminal
Published work, a patent, and a graduate R&D project. Each one has a full write-up.
Deep learning models trained on histopathology slides from one lab degrade badly on slides from another with a different staining protocol. We combine fast neural style transfer with an adversarial loss over an HRNet generator to normalize stain across labs — evaluated on whole slide images from eight different laboratories, recovering inference performance.
Documents recordings automatically by detecting visual pixel changes and semantic shifts in slides or conversation. It generates timestamps where new topics begin, splits the recording into labelled sections with summaries, and produces a transcript you can jump through by point of interest.
Built on top of the DSFD face detector to redact video shot on mobile cameras with heavy occlusion. Redaction preserves facial landmarks, so the footage stays usable as training data for other vision models without violating GDPR.
Built a website builder with a paired code/design editor from scratch in SvelteKit — canvas editing, hot reload, autosave, multipage preview. Wired GPT‑3.5 to generate context-specific templates and edit them into working features, wrote the entire component library (editor, treeview, timeline, colorinput), built a GSAP animation UI with timeline previews, and automated deploys to Vercel/Netlify with Supabase migrations via GitHub Actions.
Designer–developer tooling for sync, documentation and consistency. Integrated open-source VS Code for in-browser development with dockerized deploys and subdomain hosting; built a Figma plugin with OAuth from an Electron app to maintain design systems across accounts, generate component variants and emit Tailwind/SCSS/styled-components; shipped Figma→React and React Native codegen and an npm library to sync design tokens into codebases. Portainer, Traefik, AWS.
Took JioSE from staging to production serving 450M users, provisioning 36+ third-party bundled subscriptions including Netflix, Amazon and Hotstar. Built APIs and features for purchase notification and suspend/resume across JioTV, JioCloud and Car Connect; led a major table refactor; migrated millions of user records across Oracle environments; ran replica and prod deploys on Docker/K8s. Redis, RabbitMQ, Jest/Mocha/Chai. Patent filed from this work.
Natural-language search engine for UAV/UAM tracking — RASA for intent understanding, Elasticsearch for search, aggregation and autosuggest, with filtering and mathematical queries over flightplans, positions and subscriptions.
JEE-Advanced AIR 298 · KVPY Fellow · 10th in State Regional Maths Olympiad, qualified for INMO.
Things I built end to end. Open a case study for the detail.
Feature-parallel dev orchestrator. Several coding agents at once, each on its own git worktree and branch, with one attention queue merged across all of them — because the SDK's permission callback blocks, a queue row and a parked promise are the same object. Review, commit and ship without leaving the window.
Cloud filesystem with synchronization and block-level file encryption/decryption. Linux client for upload, download, list and sync; mobile-friendly web client for streaming video/audio/PDFs. Multi-client sync both directions with full control over the encryption schema.
Full pipeline from source to assembly — scanner, parser to AST, semantic analyser, three-address code, register transfer language and codegen. Supports arithmetic expressions, control flow and function calls.
Mobile app for splitting, tracking and settling shared expenses, paying out through any UPI app, with expense statistics.
Website builder with a paired code/design editor, AI template generation, and one-click deploys. Canvas editing, hot reload, autosave and multipage preview.
Reads the text in a photograph, translates it, erases the original by generative inpainting, and redraws the translation into the same box — matched colour, fitted font size — then speaks it. Five models chained end to end.
Point it at any pretrained torchvision model and ask what a filter has learned — deconvolution back to pixel space for what fired on a real image, and activation maximization for the pattern that would fire it hardest.
Probing what a speech recognition model keeps in its layer representations — clustering phones and spoken commands layer by layer, plus accent and speaker classification. CS 753, with a written report and deck.
Static portfolio on S3 behind CloudFront with OAC, deployed by GitHub Actions on every push to main. Vanilla JS, no build step, cache-busted assets.
Selected repositories, live from the GitHub API — @Nikhil-Chavanke-21
Open to engineering roles and research collaborations.