Anirudh Vasudevan

Top 10, Y Combinator AI Hackathon

LARK

A hierarchical multi-agent system on MCP that turns ChatGPT & Claude into action platforms.

3rd Place, ElevenHacks, ElevenLabs x Zed

Apartment 4B

A first-person lo-fi horror mystery game built end-to-end in 36 hours.

Winner, ElevenHacks, Stripe x ElevenLabs

Living Photos

Turns a single old photo into a walkable 3D scene with the voice of a loved one inside.

2nd Place, ElevenLabs x Replit Hackathon

FaceTime from Mars

Real-time voice calls with AI colonists on Mars in 2159, over an interactive 3D world.

Best Voice Agent, ElevenLabs x Firecrawl Hackathon

ScreenSense

A Chrome extension that turns voice commands into autonomous browser actions.

Production SaaS

TalkativePDF

A production SaaS that lets you chat with any PDF, using GPT-4, LangChain, and Pinecone for fast, semantic, RAG-based document Q&A.

Granted patent

CIVS

A contactless voting system, and the basis of a granted patent, for secure, hygienic, and accessible elections using speech and hand-gesture input.

LARK: Multi-Agent Orchestration MCP

A hierarchical multi-agent system built on the Model Context Protocol (MCP) that turns ChatGPT and Claude into action platforms.

A unified parent MCP orchestrator routes across three specialized child servers (media, messaging, telephony) with dynamic tool selection, making real Twilio calls and group calls, enabling cross-model AI-to-AI messaging, and handling music and YouTube, all from a phone-like widget inside chat. Built with TypeScript, React 19, and Twilio, and deployed via mcp-use.

Apartment 4B: Horror Mystery Game

A first-person lo-fi horror mystery game built end-to-end in 36 hours.

Explore procedurally generated apartments, collect voice-acted cassette tapes, and solve sequential lock puzzles under a ticking timer to unravel the mystery. Built with React Three Fiber, Three.js, TypeScript, Zustand, and Howler.js, with ElevenLabs for voice and sound. It features save state, multiple endings, and a PS1-style aesthetic.

Living Photos: Memories in 3D

Turns a single old photograph into a walkable 3D scene, with the voice of someone you loved playing softly inside.

Upload a photo and step into it: Gaussian-splat 3D rendering, consent-gated voice cloning, and procedural ambient soundscapes bring memories to life. Built with Next.js 15, React 19, TypeScript, React Three Fiber, Drizzle, Postgres, ElevenLabs, and Stripe.

FaceTime from Mars

Real-time voice conversations with AI colonists living on Mars in the year 2159.

A walkie-talkie interface lets you talk to three distinct AI characters with unique ElevenLabs voices, set against an interactive 3D Mars with radio effects and ambient space audio. Built with Next.js, React Three Fiber, FastAPI, Claude, and ElevenLabs.

ScreenSense: Voice Browser Agent

A Chrome extension that turns voice commands into autonomous browser actions.

Hold a key, speak a command, and an AI agent clicks, fills forms, and completes multi-step workflows across sites, for example “add the cheapest USB-C cable to my Amazon cart.” Powered by Claude on AWS Bedrock (multimodal vision), ElevenLabs speech, and Firecrawl, on a React + FastAPI stack.

TalkativePDF: AI PDF Chat SaaS

A production SaaS that lets you chat with any PDF, using GPT-4, LangChain, and Pinecone for fast, semantic, RAG-based document Q&A.

Built end-to-end with Next.js, TypeScript, Firebase storage, Clerk auth, and Stripe subscriptions. A RAG architecture with document chunking improved query response time by ~87%, and the app handled 500+ document uploads during internal testing.

CIVS: Contactless Voting System

A contactless voting system, and the basis of a granted patent, for secure, hygienic, and accessible elections using speech and hand-gesture input.

Built with React, Flask, and REST APIs alongside user-research-driven Figma prototypes, CIVS supports real-time interaction, ballot rectification, and accessibility for physically challenged voters. It became the basis for a granted patent (ID 202341031598).