Aditya Gupta

Framework for building multi-agent narrative simulations where world state, character actions, and knowledge live in structured, code-validated storage rather than an LLM's context window — so each character only "knows" what it's explicitly been given, actions are validated proposals rather than direct edits, and things like secrets, deception, and divergent memories can persist reliably over long sessions. It's essentially trying to fix the drift, leaks, and inconsistency that plague single-model "be the whole world" systems like AI Dungeon.

Leenim is a GUI app for building Manim math animations — code editor, node editor, and live preview in one, so you don't need to hand-write Python to make a video like 3Blue1Brown's. For humans, it removes the coding barrier. But the bigger reason it matters now is AI agents. Instead of an agent freehand-writing raw Manim/Python (fragile, error-prone, hard to run unsupervised), Leenim exposes a constrained DSL via MCP — a small set of reliable "building blocks" the agent assembles instead of guessing at code. That means agents can generate animations faster, with fewer failures, safely enough to run autonomously, and in a form the user can immediately see and tweak. It turns "AI vibe-codes Manim and hopes it compiles" into "AI reliably assembles known-good pieces" — which is exactly the kind of structured interface agentic AI needs to be trustworthy at creative, technical tasks.