Flux  ·  StudyDesk  ·  Stellar Forge  ·  de-ai  ·  Meridian  ·  Glass Ticker  ·  X  ·  YouTube  ·  Email

💡 About me: BTech CSE at Ramdeobaba University, Nagpur, second year Coordinator at Engineering India RBU, a student-led collective across 8 Nagpur colleges Building agentic systems that plan, execute, verify, and retry Real-time apps where the race conditions are the real problem Graphics and engines written from scratch, not configured The projects are small. The problems inside them are not.

Languages
Everyday
Data
Graphics
Platform

Flux: twenty-one tools behind a plan, execute, verify, retry loop. The tools were the easy part. Because streaming interleaves tool results with prose, the verifier has to judge results it has not finished receiving, and vector memory re-injects itself into later calls, so one bad extraction quietly poisons every answer after it.

StudyDesk: the Android app I use every day, which is a harder review than a stranger gives it. A photographed timetable becomes real classes through on-device OCR, so the grid parser has to survive dark backgrounds, all four rotations, and 12 px time labels that binarise into blobs. It is also built to refuse: noise cannot become a 19:20 lecture, and a rescan replaces the rows it read before instead of stacking under them. The signed APK is in the releases.

Stellar Forge: looks like a CRUD app with a 3D header, and the 3D header is the easy part. Seat claims have to be race-safe across concurrent users, so registration runs in serialisable transactions with row locking, and the live counter is one source of truth rather than whatever each client last fetched.

de-ai: a rewriter that calls no model, so the same input gives the same output every time. The cost of that determinism is that a rule bank will happily turn server into waiter, so the lexical pass is fenced off from a 600-term technical blocklist, and the pipeline re-scans its own output and reverts any swap that introduced a new tell.

Meridian: clinical triage built only from six lab practicals on real ICD-10-CM data. No ORM because none was taught, no framework for the same reason. The part I would not have guessed is that every generator bug ended up as a named regression test rather than a quiet fix.

Glass Ticker: one market ticker on five platforms - Electron on the desktop, SwiftUI on iOS, Compose on Android, and the same UI as a <glass-ticker> web component for everything else, published as npm, Python and JVM packages. The chart was the easy part. A browser cannot reach Yahoo at all, so every language binding ships a small data server, and the crumb cookie Yahoo wants does not exist in Node's fetch, so the server keeps its own. The batch quote endpoint also refuses one request in twenty, so all three servers rebuild the quote from the chart endpoint instead of drawing an empty row.

Open to internships, to collaboration, and to being told I am doing something in a stupid way.

Contribution history

Regenerated daily by .github/workflows/activity.yml and committed into this repository, so it renders with no third-party service in the path.

_{Drawn as static SVG in assets/ by me, not fetched from a badge service, so nothing on this page can break or quietly rot. The fox is a hand-drawn pixel sprite animated as a GIF, because GitHub strips inline SVG from Markdown. Both themes are designed rather than inverted, and contrast clears WCAG AA throughout.}