Hithesh Shetty.

Overview

SYSCORA is a computer-use agent, an AI that uses your computer the way you would. You type a request like "send hi to myself on WhatsApp" or "create a notes folder on my desktop with a todo list", and it does it by clicking, typing, opening apps, editing files and using a browser. Before it reports that a task is done, it checks the screen or the file system to confirm the action actually worked.

Features

• Controls apps: click, type, scroll, drag, open and close windows, change volume
• Works with files and code: find, read, create and edit files, search a codebase, read .docx, run git
• Browses the web in its own Chromium browser
• Drafts Gmail emails (you press Send yourself)
• Saves completed tasks as skills that can be replayed without calling the AI model
• Undo for some actions, an audit log, and memory of things you tell it

How it's built

• Electron desktop app connected to a local Node daemon that runs the agent loop and about 48 tools
• Reads the screen through Windows UI Automation (accessibility data), with OCR and screenshots as fallbacks
• Every tool returns a verification result, and the agent can only report success when a separate check confirms it
• Safety rules: dangerous shell commands are blocked, risky ones need your confirmation, and the terminal runs inside Windows Sandbox
• Works with your own API key: OpenAI-compatible, Anthropic, Gemini, Mistral or DeepSeek

Results

Tech stack

• Node.js
• Electron
• PowerShell
• Windows UI Automation
• OCR
• Chrome DevTools Protocol
• Windows Sandbox

Overview

You give ResearchGraph a question. It plans the research, searches the web and academic papers, pulls out claims along with exact quotes, and writes a report. Every sentence in the report links to a quote that was checked against the original source, and sentences that no source supports are removed.

Features

• Plan review: approve, edit or redo the research plan before it starts
• Parallel research: one worker per sub-question, using Tavily web search, arXiv and Semantic Scholar
• Quote checking: any quote not found word for word in its source is discarded
• Source scoring based on authority, recency, relevance and methodology
• A reviewer step that sends the run back for more research when the evidence is weak
• Dashboard with live progress, run history, evidence, sources and the final report

How it's built

• The workflow is a LangGraph state graph: Send() runs workers in parallel and interrupt() pauses for plan approval
• FastAPI backend that streams progress to the React dashboard with Server-Sent Events
• PostgreSQL with pgvector stores sources and handles passage search
• Citation numbers and the bibliography are generated in code, not written by the model
• Progress is saved after every step, so runs survive a server restart

Results

The live demo uses a mock model, so it's free to try. It's on Render's free tier and may take up to a minute to wake up.

Tech stack

• Python
• LangGraph
• FastAPI
• PostgreSQL
• pgvector
• React
• TypeScript
• Tailwind
• Docker

Output will appear here.

Overview

Create a room, share the link, and everyone in the room edits the same file at the same time. You can see each other's cursors, and when anyone presses Run, the code is executed and everyone sees the same output.

Features

• Live editing that merges changes cleanly, even when two people type on the same line
• A name, color and cursor for everyone in the room
• Runs Python, JavaScript, C++ and Java, with input (stdin) support
• Clear display of output, compile errors and timeouts
• The language setting is shared, so everyone switches together
• Monaco editor (the editor used in VS Code) with Ctrl/Cmd+Enter to run

How it's built

• React + Vite frontend with Monaco. The shared file is a Yjs document (a CRDT), which merges edits automatically
• A Node server relays changes between users over WebSockets (y-websocket), with one document per room
• Code runs through the JDoodle API via an Express proxy, so the API keys stay on the server. Runs are rate-limited per IP
• Frontend hosted on Vercel, server on Render

Results

The server is on Render's free tier, so the first load can take up to a minute.

Tech stack

• React
• Vite
• Monaco
• Yjs
• WebSockets
• Node.js
• Express
• JDoodle API
• Vitest

Overview

Chess Pal takes a screenshot of a chess board from any chess website, recognizes every piece with a trained neural network, and turns the board into a position that Stockfish can analyze. It also includes a web app where you can set up any position and play against Stockfish.

Features

• Recognizes a full board from a screenshot
• Gets the best move from Stockfish, with an option for a deeper search
• Web app to play against Stockfish from the start or from any custom position, with all chess rules
• Highlights legal moves, the last move and checks
• Small desktop tool (Tkinter) for quick best-move lookups

How it's built

• OpenCV finds the board using thresholding and contour detection, straightens it, and splits it into 64 squares
• Each square is classified by an EfficientNetB0 model fine-tuned on about 67,000 labeled square images (13 classes: 6 white pieces, 6 black pieces, empty)
• A Node/Express API sends the position to the Stockfish engine
• React frontend using chess.js for the game rules

Results

Tech stack

• Python
• TensorFlow / Keras
• OpenCV
• React
• chess.js
• Node.js
• Express
• Stockfish