Arihant Jain

Codex is the agentic copilot I built for Legistify. It sits on contracts, notices, litigation, and trademarks: Python and LangGraph beside the Node API and the React app, on AWS.

A research agent calls tools in a loop, then a separate writer answers. There is an intent gateway, a point where it can ask the person, and a guard on how long it may think. Answers stream from Claude on Bedrock and from OpenAI. Tools are isolated per tenant: search, signed copies, org metadata, live web search, and OAuth connectors for Gmail, Calendar, and Outlook.

It writes PowerPoint, Excel, HTML reports, and redlines: a Word file with real track changes and clause citations. Projects, reusable skills, and a document vault sit around that. A query guard and grounding rules keep it from inventing an id or a link.

Check live: Codex guide

Python · LangGraph · Bedrock · MCP

A backend service, not a hosted product page. An agent clones the repository, reads it, changes it, commits, and opens the pull request. Each session gets an isolated container. More than twenty can run at once.

Three specialists share the session: bugs, tests, and features. Contracting the memory made token use about five times leaner. Caching the repository cut the cold start by seventy percent.

Python · LangGraph · Docker · AWS

Two pipelines, one front door. A classifier decides whether the question needs search or a tool. Search is hybrid, dense and sparse, with contextual chunks and a cross-encoder to rerank.

The other path is a ReAct agent with more than fifteen tools, on a LangGraph state machine, with the slow work carried in the background.

Python · LangChain · LlamaIndex · LangGraph

A WhatsApp bot. A job agent, a sales agent, and a general agent share the chat, and LangGraph can hand it from one to another without starting over.

The model parses the message. The thread is held so a long task can sit and come back. More than one language.

Python · LangGraph · Twilio

A backend's own routes become a layer an agent can read. MCP-like, without standing up a separate MCP server. One core, six frameworks: FastAPI, Flask, Django, Express, Nest, Next.

The search inside it is not a borrowed library. BM25, a dictionary of twenty-one thousand terms, fuzzy match, stemming.

GitHub PyPI npm

Try now

Entangle is how I drive a Mac from a phone or a browser. Screen, keyboard, mouse, files, and chat go peer to peer. The server sets the connection up and then stays out of that traffic.

On the same network the two sides talk directly. Otherwise a public address is found, and a relay is the fallback. If a tab refreshes or the network drops, the same session comes back. Each device keeps its own key, so a new IP is not a new device.

The Mac side is a menu-bar app. It captures the screen and injects the mouse and keyboard. I put the whole service on one Linux host: the API, nginx in front of it, and the relay. It is in real use.

WebRTC · Electron · TypeScript