noesis ai

From Observation To Understanding.

Comment

Inspiration

Human curiosity about the universe, the challenges of modern astrophysics research, and the possibility of AI agents functioning as collaborative scientific researchers. What if AI agents could independently retrieve, connect, and analyze scientific information from multiple sources within a unified platform? What if an AI research system could remember previous experiments, learn from validated methods, and transfer knowledge between specialized agents?

What it does

NOESIS AI is an autonomous, multi-agent astrophysics and cosmology research platform designed to simplify and accelerate scientific investigations. It coordinates 11 specialized AI agents that collaborate on research planning, mathematical modeling, numerical simulations, data analysis, scientific validation, and report generation. The platform features a persistent scientific memory system that stores previous experiments, research findings, and reusable procedures, allowing agents to retain knowledge and continue investigations across multiple sessions. NOESIS AI performs real N-body simulations using REBOUND to study gravitational interactions and orbital stability, analyzes stellar evolution using published PARSEC stellar tracks, and explores cosmological models such as ΛCDM through numerical calculations and MCMC-based parameter estimation. It also supports astronomical observation analysis using FITS files, stellar catalogs, and light curves. By integrating autonomous AI agents, scientific computing, shared memory, and interactive visualizations, NOESIS AI provides researchers, astrophysicists, and students with a unified environment for conducting collaborative, evidence-based, and reproducible astrophysics research.

How we built it

We built NOESIS AI by integrating Agentic AI, scientific computing, persistent memory, and full-stack web development into a unified astrophysics research platform. We developed the backend using Python and FastAPI and implemented a multi-agent orchestration system consisting of 11 specialized agents responsible for research planning, mathematical reasoning, astrophysical simulations, observational analysis, scientific validation, and report generation. To enable long-term knowledge retention and collaboration, we designed three persistent memory systems—episodic, semantic, and procedural—using PostgreSQL, pgvector, and FalkorDB to store experiments, retrieve scientific knowledge, and connect related research findings through a knowledge graph. We integrated scientific computing libraries, including REBOUND, Astropy, NumPy, SciPy, and emcee, to perform real N-body simulations, stellar evolution analysis, cosmological calculations, and statistical parameter estimation. We developed an interactive frontend using React, TypeScript, Vite, and Tailwind CSS, featuring scientific dashboards, simulation visualizations, an agent coordination interface, and a memory explorer. We also implemented durable research workflows, experiment validation, session recovery, and reproducible report generation. Finally, we containerized the application using Docker, added automated testing, and configured deployment through Vercel, creating an integrated environment where specialized AI agents can collaborate, execute scientific computations, and preserve research knowledge across sessions.

Challenges we ran into

One of the biggest challenges was coordinating 11 specialized AI agents while ensuring reliable communication, task dependencies, shared state, and recovery from interrupted research workflows. Implementing persistent scientific memory was equally complex, as we needed to manage episodic, semantic, and procedural knowledge across multiple sessions while preventing conflicting updates and maintaining data consistency. Integrating FalkorDB introduced additional challenges involving knowledge graph persistence, tenant isolation, and fault tolerance. We also faced difficulties integrating real astrophysics computation libraries such as REBOUND, Astropy, SciPy, and emcee, particularly when ensuring numerical accuracy, physical consistency, and reproducibility. Stellar evolution modeling required careful handling of published PARSEC tracks, interpolation, and scientific assumptions. Another significant challenge was connecting computationally intensive backend operations to an interactive frontend while providing real-time agent activity updates and scientific visualizations. During deployment and testing, we encountered Docker environment instability, database configuration complexities, and issues maintaining reliable communication between services. We addressed these challenges through durable orchestration, scientific validation checks, version-controlled memory, fault-tolerant graph integration, automated testing, and reproducible experimental workflows.

Accomplishments that we're proud of

  1. Multi-Agent Orchestration: Developed a coordinated system of 11 specialized research agents with dependency-aware task execution, shared state management, and automated scientific workflows.
  2. Persistent Multi-Layer Memory Architecture: Implemented episodic, semantic, and procedural memory with version control, enabling cross-session knowledge retention and reuse of validated research procedures.
  3. Graph-Based Knowledge Retrieval: Integrated FalkorDB to create a tenant-scoped scientific knowledge graph connecting research sessions, experiments, observations, procedures, and validation results.

What we learned

  • Multi-Agent Orchestration
  • Persistent Multi-Layer Memory Architecture
  • Graph-Based Knowledge Retrieval
  • Scientific Computing Integration
  • Stellar Evolution Modeling
  • Interactive Research Visualization

What's next for noesis ai

  • Enhance our 11-agent architecture with dynamic task planning, hypothesis generation, adaptive experimentation, and self-correcting research workflows using advanced LLM reasoning.
  • Extend our episodic, semantic, and procedural memory systems with knowledge consolidation, intelligent retrieval, contradiction detection, and evidence-based memory updates.
  • Explore reinforcement learning and feedback-driven optimization to improve agent coordination, task allocation, and scientific workflow efficiency.
  • Implement researcher approval checkpoints, uncertainty quantification, independent verification, and transparent scientific reasoning to improve reliability.

Built With

Share this project:

Updates

Submission history