Second-year CSE · Manipal University Jaipur
Shreyas
Fegade
Raw signal into legible intelligence.
I build software I'd actually use myself — six products unified by one idea: tools that turn messy raw signal into something you can read, rendered through custom-built visualization, never dropped-in libraries.
Cortex
An AI-native developer workspace in one window.
Terminal, browser, file explorer, tasks and a multi-provider AI assistant collapsed into a single keyboard-driven Electron shell. A passive telemetry layer records the workday to local SQLite and feeds it straight back to the assistant as selectable context.
- Electron 35
- React 19
- TypeScript
- 120 typed IPC channels
- node-pty
- SQLite (WAL)
A two-process workspace resolving into a context engine
REGIME
Four hidden market regimes, classified and stress-tested.
A 4-state Gaussian Hidden Markov Model reads return, volatility, volume and momentum to infer the unobserved regime a market is in — then measures the forward edge and runs a walk-forward, out-of-sample backtest. Built India-first for NSE/BSE, rendered through a hand-drawn HTML5 Canvas engine.
- Python
- Gaussian HMM
- FastAPI
- Canvas engine
- GSAP
- yfinance
A four-state HMM lattice over a candlestick stream
Knowledge Mapper
A dense PDF, rendered as a map of how its ideas connect.
A three-stage LLM pipeline — global understanding, hierarchical concept extraction, topology inference — turns a document into a force-directed concept graph with typed, justified relationships. The renderer is a from-scratch 2D canvas engine with its own force simulation and spatial-hash hit testing. No graph library.
- Next.js 15
- React 19
- FastAPI
- Canvas force sim
- Spatial hash
- SSE
- SQLite
A force-directed concept graph settling into a galaxy
Forensics
See what your files are hiding — entirely in the browser.
Drop any file to expose EXIF metadata, GPS leaks, true file type, byte-level Shannon entropy, embedded strings and appended steganographic payloads. A hand-rolled binary EXIF/TIFF parser and entropy chart, with zero uploads and zero dependencies — nothing ever leaves your machine.
- Vanilla JS
- ArrayBuffer + DataView
- Canvas
- Shannon entropy
- 0 dependencies
A 256-bar entropy histogram beneath a scanning reticle
Chronicle
Passive focus intelligence, measured as entropy.
A Windows tray daemon samples the foreground window and turns attention into a single honest number — a Focus Score built from normalized Shannon entropy — rendered through a hand-built D3 Focus Stream where turbulence is the entropy itself. Native Win32 via ctypes, WAL-mode SQLite, everything local.
- Python
- Win32 ctypes
- Shannon entropy
- FastAPI
- SQLite (WAL)
- D3.js
A focus ribbon fraying from calm into turbulence
Unbored
The decision-paralysis killer — one confident pick.
Pick a mood, tap once, get one thing to watch right now. A weighted scoring engine ranks candidates from TMDB and AniList across genre, keyword, mood, runtime and diversity, then a pluggable LLM writes one honest 'Why now?' sentence. Procedural poster art for artless titles; 302 tests.
- React 19
- TypeScript
- FastAPI
- TMDB + AniList
- Gemini / pluggable LLM
A field of candidates collapsing to a single confident pick
Who this is
Shreyas Fegade — building tools that read the signal.
Second-year CSE at Manipal University Jaipur, 2025–2029, based in Jaipur, India. Six projects, one thesis: Tools that turn messy raw signal into legible intelligence — every visualization hand-built, never a dropped-in library.
Custom visualization, always
Every chart, graph and field is drawn from the data by hand — no Cytoscape, no charting library. The rendering is the product, not an afterthought.
Signal in, legibility out
EXIF bytes, market regimes, attention entropy, a dense PDF — each project takes a messy raw signal and makes it something a person can actually read.
Treated as real products
Real loading, empty and error states. Keyboard-first. Privacy by default. Built to be used, not just demoed.
Open for internship opportunities