Matrix

Inspiration

Market research is hard.

Understanding how a rumor, bill, or breaking news event will ripple through a population is extremely difficult. Traditional Markov chains fail to capture the inconsistent nature of intelligent subjects. Polling is slow, expensive, and surface-level. Focus groups can't capture the emergent dynamics of thousands of interconnected people with different backgrounds, relationships, and biases. And once a campaign or policy launches, it's too late to observe the spread; we only see the aftermath.

Currently, no tool exists that enables one to accurately model a realistic, human social network.

We changed that.

With Matrix, one can run both short and long-term information diffusion simulations across demographically diverse agents, and also even interview those agents to gain low-level insights to the distribution. By creating digital twins grounded in truth to accurately study a population parameter without the hassle of space, time, and ethical constraints, we enable researchers to find signals amidst the noise of human nature.

What it does

Orchestration Layer: Given a user’s query, Matrix leverages Exa to find the latest real-world data for an agentic Planner model. This Planner, powered by either the OpenAI API or a Modal-hosted inference session, then leverages various sources such as census data to generate realistic sample distributions of agentic digital clones. Each agent is grounded in truth through strict prompting & fine-tuning, aligning itself with its given parameters.

Simulation Layer: Matrix leverages numerous inferencing sessions to allow our sample population distribution to interact just as individuals would in the real world. Thus, we allow researchers to study the propagation of signals directly in the sample distribution, controlling for confounding variables and resulting in more accurate insights in their studies. Here, we maintain model accuracy & realism through a mix of LLM selection strategies, the use of Supermemory, and a breadth-first approach.

Presentation Layer: To bundle the large amount of data Matrix so quickly acquired, we compile our findings into PDF & LaTeX documents for the end user. Additionally, users can study notes & summaries of each agent’s day-to-day interactions or opt to directly interview agent participants, allowing for even more granular insights to signal propagation. All of these are enabled through the use of the ElevenLabs API (for voice) and OpenAI API (for image/video).

How we built it

Our React + Three.js frontend is connected to our Python backend via FastAPI. Models, such as the Deepseek R1 32b we use for the Planner, are hosted on Modal. They are enhanced in capability via the Exa API for web search (precision) and Supermemory for tracking (recall). Interviews are enabled through the use of ElevenLabs and LiveKit, and test-to-speech is powered by OpenAI’s Whisper.

Agent Logic

Matrix uses a tiered multi-agent architecture where each agent is tuned with specific parameters and tools to fulfill its role in the social ecosystem.

The Planner Agent utilizes Exa to retrieve real-world data, such as census reports or local public news forums, and generates a sample population datafile. This datafile contains granular, agent system-level instructions & personality seeds such as age, gender, beliefs, and more.

The Social Network Graph implements a Breadth-First approach of information propagation, similar to the real-world, to model realistic connection topologies and route messages between neighboring agents.

Matrix’s Simulation Agents execute N-day parallel cycles, where agents process neighbor messages and update their internal belief statements without sequential bottlenecks. This is then reciprocated, allowing for meaningful conversation between independent personalities.

Our Isolated Context Routing enforces a strict architecture where workers only receive data from direct neighbors, preventing context pollution and keeping the simulation clean.

Our Persistent Evolution system captures daily snapshots of interactions and belief shifts via Supermemory.

Challenges we ran into

Getting agent personas to stay distinct across a 5-day simulation without drifting toward generic consensus responses
Cold start latency on Modal containers under burst load from parallel simulation turns
Managing the 26KB context ceiling for planner prompts when users upload large documents
LiveAvatar session lifecycle management — tokens expire mid-conversation and need graceful refresh
Keeping the social graph coherent after LLM-generated connections, agents referencing IDs that don't exist, self-loops, and reciprocity imbalances required extensive validation logic

Accomplishments that we're proud of

Successfully leveraging Modal to the fullest extent possible–we were able to host and query hundreds of agents at any given moment, effectively allowing us to create digital twins of entire towns and cities.
Integrating Supermemory to agents, transforming them from simple transfer-based ‘next-token predictors’ to almost real, concrete individuals
Developing an entire interview pipeline, allowing users to communicate with virtual models just as they might do via Discord or Zoom

Visit site here: https://www.ireallywantmetaraybanglasses.tech/