Projects

Project

Category

Context

Year

Roentgen

Machine learning

ISRO BAH 2026

2026

RoentgenTeam Navya

An offline, auditable decision-support system for solar flares, built on data from India's Aditya-L1 spacecraft: two independent X-ray detectors that call flares as they happen, and a 24-hour forecaster for every active region on the Sun.

The brief

Problem statement 15 of ISRO's Bharatiya Antariksh Hackathon 2026: forecast and nowcast solar flares from Aditya-L1's soft and hard X-ray instruments, SoLEXS and HEL1OS. Flares disrupt satellites, GPS, radio and power grids, so the system answers three questions for an operator: is a flare happening now, and how strong; will one happen in the next day; and what has happened, as a catalogue.

Data

  • SoLEXS soft X-ray and HEL1OS hard X-ray Level-1 data from ISRO's PRADAN archive, merged into one causal 1-minute UTC frame with a mask for blackouts and gaps.
  • SDO/HMI SHARP magnetograms: 18 magnetic parameters per active region, from 2010 to 2026, plus the geometry of each region's polarity-inversion line.
  • GOES and HEK flare catalogues, used only to validate and calibrate, never as a training input for the X-ray models.
  • About 4.26 million fused rows: 3.83M for training, 194K for validation and a sealed 236K-row blind set.

How it works

  • Two nowcasters, one per instrument, that share no features, labels or masks: each reads only its own channel's level, slope, acceleration and rolling statistics, and finds flares as excess over that channel's own background. They meet in one engine, and every flare goes into a master catalogue tagged soft, hard or both.
  • The soft channel reports GOES classes; the hard channel reports burst tiers relative to its background, on purpose not GOES letters. The soft–hard (Neupert) relationship is measured per flare afterwards, never assumed.
  • A 24-hour forecaster of XGBoost with a precision arm and a recall arm, blended, over SHARP keywords, magnetogram geometry, flare history and X-ray context. Isotonic calibration, then two alert tiers: WATCH (catch first) and WARNING (precision first), debounced so they don't flicker.
  • The models are small on purpose: about 18K parameters for the forecaster, against a 1.8M-parameter deep model that lost the honest test.

Making the numbers honest

  • An early score of 0.994 AUC turned out to be circular (labels from the same channel the model read). It was thrown out and every feature made causal.
  • Folds never share an active region between training and validation, run forward in time, with a 72-hour purge between them. A label-shuffle placebo falls to chance, and a truncation test finds no feature that sees the future.
  • A sealed blind set, scored at most once, with every scoring logged. Every published number has a receipt: a SHA-256 integrity manifest and per-seed predictions.
  • Gains that turned out to be leakage were withdrawn, and the docs say plainly what the system is not: a deployed real-time service, or a ≥M forecast that beats climatology.

Results

  • Hard X-ray nowcast: 95.7% of held-out flares caught at 0.29 false episodes a day, about 100× fewer false alarms than untuned.
  • 24-hour forecast across five folds: ROC-AUC 0.964 and transfer TSS 0.793 on validation; every X-class flare in the window caught at the wide gate.
  • SWAN-SF benchmark: 0.808 mean TSS over three seeds, against 0.748 published for GCTAF (with different preprocessing, so no leaderboard claim).
  • The sealed blind set, spent once: PR-AUC 0.515, TSS 0.741.
  • Inference in 0.36 ms per active region on a CPU, against SHARP's 12-minute cadence. 91 tests, CI, and a live week shadowing GOES that agreed with NOAA 16 times out of 16.

The dashboard

A web dashboard (Next.js, live as a demo) replays real archive days, such as the X9.0 flare of 3 October 2024: both light curves with GOES events shaded, the nowcast's state and class probabilities, the system's health, a catalogue of recent events with how early each was called, and tabs for signals, the model and a simulation. An offline Streamlit version takes uploaded SoLEXS and HEL1OS files.

Event
ISRO Bharatiya Antariksh Hackathon 2026, top 34
Team
Team Navya
Timeline
June – August 2026
Status
Prototype; demo replays archive data
Stack

PyTorch

XGBoost

astropy

Next.js

Streamlit

XceliQ

SaaS

Euclideum

2026

A cognitive practice and assessment platform at Euclideum Solutions, where I'm a software developer intern. I engineered its scoring engine and rebuilt its 36 games on one shared UI kit; separately, I made the company's own site several times faster.

Scoring engine

XceliQ measures people across eight dimensions (memory, reasoning, attention, planning, spatial, decision making, adaptability and workplace) through 36 short games. The engine turns raw trials into scores in stages:

  • A quality gate that down-weights rushed or repeated sessions.
  • Difficulty measured from play rather than assumed, and an expected-performance model, so a hard round counts for more.
  • Game-specific metrics where they mean something: d′, search slopes, Weber fractions, planning waste, mirror confusion, a Rescorla-Wagner fit. Memory and decision games never score speed.
  • Median/MAD normalisation to 0–100 against a population norms table refreshed nightly, shrinkage for thin data, and split-half reliability (Spearman-Brown) behind a confidence label of High, Building or Low.
  • One game was taken out of scoring after finding its answer key was shuffled separately from its replies, so its score measured luck.

Games and product

  • All 36 game renderers rebuilt on a shared React UI kit, driven by JSON trial definitions, each verified by its own round-trip test.
  • A cognitive profile page: a radar chart with no layout shift and keyboard-reachable axes, ranked bars and a trajectory over time.
  • The assessment runner, with timeout warnings and a confirmed hand-in, and the admin assessment builder.

The company site

I moved euclideum.com from Next.js 14 to 16 and then went after its speed, measured as the median of five runs on a 4× throttled CPU:

  • Mobile Lighthouse from 38 to 52, with the preloader rebuilt in pure CSS so the largest paint no longer waits for hydration.
  • Main-thread blocking on the home page from 8.4s to 1.5s; on the School ERP page from 6.3s to 0.65s. Scrolling went from about 8 fps to 21–34.
  • The biggest win was content-visibility on sections below the fold; the logo rail went from four copies to two, the ribbon canvas capped at 30 fps, and dynamic imports tried and reverted when they made idle frame rates worse.
  • Six copies of a reveal-on-scroll hook merged into one, which fixed sections that stayed blank (14 stuck elements on the home page, now none).
  • 77 security advisories patched, and lint warnings from 6,474 to 298.
Role
Software developer intern
Timeline
June 2026 – present
Team
Euclideum's engineering team
Stack

React

Next.js

Fastify

Postgres

Turborepo

Aditya-L1 CME Detector

Machine learning

ISRO BAH 2026

2026

A research prototype that gives, for each few minutes of solar-wind data from Aditya-L1, the probability that a coronal mass ejection is passing the spacecraft.

Why

Coronal mass ejections drive geomagnetic storms, like the G5 storm of May 2024. Aditya-L1's ASPEX-SWIS instrument samples the solar wind at L1, ahead of Earth, but from a single point and in the noisy conditions of solar maximum, where false positives are easy.

How it works

  • Level-2 CDF plasma data, with fill values removed, short gaps bridged and shocks kept sharp by Savitzky–Golay smoothing.
  • Eight physics features: solar-wind speed, proton density, temperature, the helium-to-hydrogen ratio, a plasma-beta proxy, and the rates of change of speed and density.
  • A causal temporal convolutional network: six dilated residual blocks (dilation 1 to 32) over 128-step windows, a receptive field of 253 steps, about 152K parameters.
  • CMEs are 3.2% of the 44,103 windows, so a weighted loss and a tuned threshold; alerts escalate as WATCH, WARNING and ALERT, and the model exports to ONNX.

Results and limits

On re-run, test ROC-AUC is 0.87 with an F1 of 0.25, catching about half to three fifths of CME windows. The repository is open about what still needs fixing: windows overlap across the random train/test split, the temperature input is constant, and the published checkpoints don't load with the shipped loader. Treat it as a research prototype, not a forecaster.

Event
ISRO Bharatiya Antariksh Hackathon 2026
Timeline
May 2026
Data
ASPEX-SWIS Level 2, from ISSDC
License
MIT
Stack

PyTorch

SciPy

ONNX

Streamlit

SVAN M2 Navigation

Robotics

Research

2026

Natural-language control and a real-time vision pipeline for the xTerra SVAN M2 quadruped robot: you tell it what to do, show it a gesture, or walk and it follows, and it stops for what's in its way.

The patent

A pending patent covers natural-language navigation: an LLM acts as a hierarchical task planner, breaking a goal down into the robot's basic actions and re-planning as things change, with perception from an Intel RealSense depth camera, SLAM and semantic segmentation, a vision-language model to find what you refer to, and processing at the edge on an NVIDIA Jetson.

The vision pipeline

The open-source half runs entirely on a Jetson Nano; the cloud was ruled out for latency, connectivity and safety. Each frame goes through one shared YOLOv8n pass, split into people and obstacles:

  • Gestures in 3D: joint angles and a palm coordinate frame from MediaPipe, so a gesture reads the same however the hand is turned (image-space rules fail the moment it rotates). Thirteen gestures, eleven of them commands, steadied by a majority vote with hysteresis.
  • Person following: the most confident, largest person, steered with a dead-band and slowed in steps as they get close, down to a stop.
  • Obstacles in 18+ classes, zoned near or medium by how much of the frame they fill, a stand-in until depth comes from the RealSense.
  • A priority arbiter decides what wins (emergency stop over stop over following), with a cooldown that a stop always bypasses, and commands go to the robot's ROS controller over HTTP with a heartbeat.

Making it fast

On the Jetson the detector is a TensorRT FP16 engine built on the device, at 416px, with gesture recognition every second frame and a camera thread that drops stale frames; settings switch automatically when it detects a Jetson. In field use it runs at around 16–20 fps, against 5–10 without TensorRT.

What's next

Distance still comes from box size rather than depth, and control is reactive. The project's own report lists its known defects and risks, such as a stop that can be masked while following, to fix before more autonomy.

Status
Patent pending
Hardware
xTerra SVAN M2, NVIDIA Jetson
Timeline
2026
License
MIT (vision pipeline)
Stack

YOLOv8

MediaPipe

TensorRT

ROS

OpenCV

MyRAIT

Web app

RAIT students

2026

A faster way into MyDY, the university's Moodle, for RAIT students: subjects, documents, attendance, assignments, grades and an AI-read syllabus in one place. MyDY regularly takes 10–15 seconds just to show a dashboard.

How it works

  • You sign in with your own MyDY account; the server logs in, takes the session apart for its keys, and gets a Moodle web-service token. The app then keeps its own signed session, with silent re-authentication when MyDY's expires.
  • Hybrid access: Moodle's web services where they exist (courses, contents, assignments, submissions, grades, messages, and the university's own attendance functions), scraping its pages where they don't.
  • The results portal, reverse-engineered from its own JavaScript, for importing results into a CGPA/SGPA calculator, keeping the latest declared result where a semester had re-exams.
  • A typed API of 22 routes over all of it, with caching tuned per kind of data, from 5 minutes for semesters to an hour for documents.

The syllabus pipeline

Each course's syllabus PDF is read by Mistral OCR, checked against the course's title, and turned by a language model into structured modules, exam scheme, credits and books, stored in Supabase and processed in the background. There's a daily refresh limit, a manual upload fallback, and an AI overview you can open in ChatGPT, Gemini, Claude, Perplexity or Le Chat.

Built for a slow upstream

  • It tells MyDY being slow from MyDY being down (gateway errors mean down; timeouts mean slow) and says which.
  • Generous timeouts for the slowest pages, and file links accepted only from MyDY itself.
  • Installable as an app, with haptics and UI sounds, and preferences such as hiding empty weeks or reducing blur.
Users
200+ RAIT students, daily
Timeline
March 2026 – present
Role
Solo: design and engineering
Stack

Next.js

TanStack Query

Supabase

Mistral OCR

cheerio

Scriptorium

AI, Desktop app

Personal

2025

Document intelligence that runs on your own machine. It started as a front end for Mistral's OCR, before Mistral had a UI of its own for it, and is growing into a desktop suite that runs open vision-language models offline.

The web app

The live beta (OCR Ready) reads PDFs and images with Mistral OCR or Gemini, with your own API keys kept on your side, a PDF viewer, and a reformatting step that cleans up what comes back.

The desktop suite

In progress: an Electron app that routes each model to the runtime that suits it, all offline:

  • A catalogue of 29 models: OCR models such as Qwen3-VL, DeepSeek-OCR, olmOCR-2, Nanonets-OCR2, PaddleOCR-VL, InternVL3, Granite-Docling and TrOCR, and small text models for cleanup.
  • Run through Ollama, a bundled llama.cpp server for GGUF models, or transformers.js on the CPU.
  • OCR fusion (ROVER): several models read the same page, their transcripts are aligned word by word, and they vote.
  • Prompts tuned per model, because some lose quality without their exact prompt, and instruct variants where thinking ones return empty pages.
  • A LaTeX editor, an AI chat panel, page arrangement and an annotation whiteboard.

The vectorizer

Raster to SVG: upscale and denoise, snap to the image's dominant palette, trace, then a geometric refiner that swaps traced paths for exact circles, rectangles, polygons and arcs wherever they fit within tolerance, and SVGO. Photos are detected and skip the palette snap.

Timeline
November 2025 – present
Status
Web beta live; desktop in progress
Role
Solo
Stack

Next.js

Electron

Ollama

llama.cpp

transformers.js

Routix

Optimisation

SIH 2025

2025

A train throughput optimiser for Smart India Hackathon 2025, co-developed as a prototype.

What it does

It models a railway section from operational data (congestion, dwell times and delays) and tracks how efficiently the network runs through a set of KPIs, as a basis for deciding which train goes first.

Event
Smart India Hackathon 2025
Role
Co-developer
Status
Prototype
Stack

Optimisation

Data modelling

Rudhirsetu

CMS migration

Rudhirsetu NGO

2025

Rudhirsetu Seva Sanstha

The website of Rudhirsetu Seva Sanstha, an NGO in Panvel that has run blood donation drives, eye check-ups, cancer awareness and thalassemia support since 2010.

What I did

  • Moved the site onto Sanity CMS, with its content (camps, gallery, initiatives) modelled as structured schemas the team can edit themselves.
  • Standardised SEO and Open Graph metadata across pages, so shared links show a proper preview.
  • Checked that nothing was lost in the migration.
  • Eliminated the layout shift, so the page no longer jumps as it loads.
Role
Software developer intern
Timeline
October 2025
Stack

Next.js

Sanity

SEO

Music Motion

Reverse engineering

Personal

2025

Animated album covers on your lock screen and desktop. Paste an Apple Music or Spotify link and it finds the song's motion artwork; the Android and Windows apps then play it as a live wallpaper for whatever you're listening to.

Reverse engineering

  • Apple Music: the server reads the token Apple's own web player embeds in its scripts, checks it, and keeps it until just before it expires. The animated cover is an HLS stream on the album's editorial video, tried tall, then square.
  • Spotify Canvas: a web-player token, then a hand-written protobuf request to Spotify's canvas service, with the response decoded by hand.
  • Matching a song to its album by fuzzy similarity of title and artist, with a minimum score and no fallback to the first result, and a search through the artist's albums when the song itself has none.
  • Animated covers can be downloaded as MP4, stitched from the stream's segments.

Liquid Glass

Its interface was one of the first implementations of Apple's Liquid Glass on the web: an SVG filter whose displacement maps are computed pixel by pixel from the shape of a bezel, the glass's thickness and its refractive index, with a specular highlight, animated with springs.

The apps

  • Android (Kotlin, Jetpack Compose): reads what's playing from notifications and plays the cover in a wallpaper service with ExoPlayer. Reworked for battery: no wake lock or foreground service, paused whenever the wallpaper isn't visible, results cached per track.
  • Windows (.NET 8): a tray app that reads Windows' Now Playing and feeds a Wallpaper Engine wallpaper over a local WebSocket, crossfading between covers.
  • On the web: charts of the most-played animated albums by country.
Timeline
October 2025 – July 2026
Platforms
Web, Android, Windows
Role
Solo
Stack

Next.js

Kotlin

Media3

.NET

hls.js

Virtual Try-On Engine

Diffusion inference

Personal

2025

FASHN VTON v1.5, a virtual try-on model, made to run on a consumer graphics card.

What changed

  • Efficient diffusion inference and caching, and a segmentation-free MMDiT pipeline.
  • Four times faster: from 40 to 10 seconds per image on an RTX 5070 Ti.
  • Memory from about 19.5 GB of VRAM down to 2.9 GB, so it fits on ordinary GPUs.
  • Served over FastAPI, with sessions that try on several garments.
Hardware
RTX 5070 Ti
Stack

PyTorch

Diffusion

FastAPI

Technovate '25

Shaders, UI

RAIT

2025

D Y Patil University, Ramrao Adik Institute of Technology

The platform for Technovate '25, RAIT's technical festival, and its hackathon.

My part

Optimising its shaders, and smaller improvements across the interface.

Role
Shader optimisation, UI
Stack

WebGL

Shaders

Vite

Blockchain KYC

Blockchain

Personal

2025

Identity verification tied to a crypto wallet: verify once, and the proof travels with the wallet. It supports over 1,000 wallets.

The flow

  • Connect a wallet.
  • Upload identity documents (Aadhaar and PAN required; passport or voter ID optional), checked by AI.
  • A live selfie, to prove a person is present.
  • The result is stored on IPFS and sealed with a hash.
Stack

Next.js

Web3

IPFS

Traversal Playground

Visualiser

Personal

2025

An interactive visualiser for graph traversal: build a graph, then step through a traversal one move at a time and inspect its state as it runs.

Stack

React Flow

Zustand

TypeScript

Hacktoberfest '24

Event platform

GDG RAIT

2024

Google Developer Groups, Ramrao Adik Institute of Technology

The platform for GDG RAIT's Hacktoberfest '24, a two-day event celebrating open-source contribution, used throughout it.

Context

As GDG RAIT's technical lead I led Hacktoberfest '24 itself, along with hackathons and workshops, and built the sites and dashboards for them, used by over 1,000 students.

Role
Technical lead, GDG RAIT
Timeline
September – October 2024
Stack

Next.js

TypeScript

RAIT Railway Concession Software

Internal tool

RAIT

2024

D Y Patil University, Ramrao Adik Institute of Technology

The college's system for railway concessions: students apply for the form behind a discounted train pass, and the office verifies, prints and exports the applications.

How it works

  • Students register and fill in the form (roll number, year, branch, home station, class and duration); their age is worked out from their date of birth, a second application is caught, and a confirmation email goes out.
  • The office verifies each application, prints the concession form, and exports everything to Excel.

What I did

As part of the technical team, improved how receipts are generated and line up when printed, and optimised its queries for the load of the whole college using it.

Role
Technical team member
Based on
Railway-ConCession (MIT, Zaid Parkar)
Stack

PHP

MySQL

Bootstrap

RAIT Semester N

Student resource

RAIT students

2026

AI & ML Semester 7, unofficial

Everything my class needed for the sixth semester of B.Tech AI & ML, in one searchable place.

What's in it

  • Six subjects (Advanced ML, Deep Learning, Big Data Analytics, NLP, Design of Experiments, Generative AI) with curated question banks for each exam, module notes and solutions.
  • Every lab: AML, DL, BDA, MLOps, NLP and DoE experiments.
  • The syllabus, academic calendar, timetables and exam schedules.
  • Ask AI on any page, opening it in ChatGPT, Gemini, Claude or Perplexity, with OCR'd notes and papers as context.

How it's built

Fumadocs on Next.js, 177 pages of content in MDX with maths and diagrams; curated question banks are gated to signed-in members with Supabase. It carried on as RAIT Semester 7.

Timeline
February – July 2026
Role
Solo
Stack

Next.js

Fumadocs

MDX

Supabase

The Indian Pantry

E-commerce

The Indian Pantry

2025

A headless online store for The Indian Pantry, a brand of roasted makhana and Indian-spiced snacks, on Next.js with Shopify behind it.

Shopify, headless

  • Products, collections and carts through Shopify's Storefront API, retried on patchy mobile connections, and errors surfaced rather than showing an empty store.
  • The cart kept across visits, and checkout handed to Shopify's own, pointed back at the real site.

The store

  • Product pages with photo and video galleries, similar products, categories and festive combos.
  • A slide-out cart with a sticky checkout, and one free-delivery threshold (₹299) shared by the banner and the cart's progress bar.
Role
Built the store; the client updates content
Timeline
September 2025 – present
Stack

Next.js

Shopify

Tailwind CSS

Safe Pro

Website

Safe Pro

2026

The website and product catalogue of Safe Pro Fire Services, a fire-safety manufacturer: extinguishers, hydrant and suppression systems, certificates and branches.

What's in it

  • Products by category, drawn from the company's own backend, with spec tables, quote requests and enquiries.
  • Blogs in MDX, support through WhatsApp and Zoho Desk.

Search and answer engines

  • A move to clean slug-based URLs, with 49 permanent redirects so nothing broke.
  • Structured data for the organisation, products, articles and breadcrumbs, a dynamic sitemap, and IndexNow.
  • Crawlers of search and AI answer engines explicitly allowed, and a knowledge centre of answer-first articles in progress.
Timeline
March 2026 – present
Stack

Next.js

MDX

GSAP

Canvas

Shaders, DOM

Personal

2026

A test of shaders and the DOM working together on one page.

Stack

WebGL

Shaders

DiffuseFit

AI

Hackathon

2025

Virtual try-on with diffusion: give it a photo of a person and of a garment, and it shows the person wearing it. Built at RAIT with a teammate, it won its hackathon.

How it works

Built on IDM-VTON, a Stable Diffusion XL try-on model. The person's pose (OpenPose), body parts (human parsing) and body surface (DensePose) mark where the garment goes; the garment is encoded separately and the model paints it on, at 768×1024. Images of other shapes are cropped and pasted back afterwards.

Timeline
August 2025
Stack

PyTorch

Diffusers

Gradio

Next.js

Crypto4College

Blockchain

Hackathon

2025

Where campus meets blockchain: crypto, made for college life. A hackathon finalist.

Stack

Next.js

Web3

Tangerine Convert

Tool

Personal

2025

File conversion that never leaves your device: FFmpeg runs in the browser, so nothing is uploaded. Its dialogs morph smoothly from one step to the next, and it does simple calculations too.

Stack

Next.js

FFmpeg

Motion

RefractUI

UI kit

Personal

2025

Landing-page building blocks, led by glowing bento grids whose layout you can adjust. The code is yours to copy and fit to your stack.

Stack

Next.js

Tailwind CSS

SaaS Template

Template

Personal

2025

A template for an AI product's landing page: the pitch, the teams that trust it, the problem it solves and how.

Stack

Next.js

Tailwind CSS

Brand Landing Page

Template

Personal

2024

A landing page template for a design engineering studio, with lettering that animates in letter by letter.

Stack

Next.js

Framer Motion

Tailwind CSS

The Wall

Website

RAIT

2023

The site of The Wall, RAIT's literary club: its blog, reviews, magazines, members and a memories lane, on Sanity. Now deprecated, and many of its images are missing.

Stack

Next.js

Sanity

Home