Inspiration
What happens when a bank loses millions of dollars? How does that affect businesses, jobs, and the wider economy? And what can be done to prevent a crisis from spreading?
Financial crises are interconnected, but their consequences are often difficult to understand. A bank's losses can restrict lending, leaving businesses struggling to fund payroll and operations. Those disruptions can spread to households and other financial institutions.
I built Econova to make these relationships understandable through interactive simulations and AI-powered investigations.
What it does
Econova is an interactive economic simulation where users create a virtual economy of banks, businesses, and households, introduce financial shocks, and explore the consequences.
Users can:
Simulate financial crises: Apply a hypothetical investment loss to a bank and observe the effects on lending, employment, production, and defaults.
Explore alternate realities: Compare different bank shocks from the same initial economic state.
Compare rescue strategies: Use Fork Reality to test no intervention, emergency bank lending, and direct household support.
Watch AI agents collaborate: An investigator gathers financial evidence, a policy planner evaluates interventions, and an independent supervisor reviews recommendations.
Negotiate credit and purchases: Business, banking, buyer, and supplier agents evaluate financing and procurement decisions.
Inspect evidence: Explore recorded tool calls, agent handoffs, simulation results, and decision explanations.
Maintain human oversight: Financial actions require explicit approval and cannot proceed when required reviews fail.
Econova is designed as an educational investigation tool, not a real-world economic forecasting service.
How we built it
NVIDIA Nemotron + Nebius Token Factory
Econova uses NVIDIA NEconova is an interactive economic simulation where users create a virtual economy of banks, businesses, and households, introduce financial shocks, and explore the consequences.
Users can:
Simulate financial crises: Apply a hypothetical investment loss to a bank and observe the effects on lending, employment, production, and defaults.
Explore alternate realities: Compare different bank shocks from the same initial economic state.
Compare rescue strategies: Use Fork Reality to test no intervention, emergency bank lending, and direct household support.
Watch AI agents collaborate: An investigator gathers financial evidence, a policy planner evaluates interventions, and an independent supervisor reviews recommendations.
Negotiate credit and purchases: Business, banking, buyer, and supplier agents evaluate financing and procurement decisions.
Inspect evidence: Explore recorded tool calls, agent handoffs, simulation results, and decision explanations.
Maintain human oversight: Financial actions require explicit approval and cannot proceed when required reviews fail.
Econova is designed as an educational investigation tool, not a real-world economic forecasting service.emotron 3 Nano (nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B) through Nebius Token Factory to power its financial AI agents.
The agents use structured function calling to gather evidence, inspect economic conditions, test interventions, negotiate offers, and independently review proposed decisions.
The central workflow consists of:
Investigator Agent: Collects relevant financial evidence through simulation tools.
Policy Planner Agent: Tests permitted intervention strategies and proposes an evidence-backed response.
Supervisor Agent: Independently evaluates the recommendation and can hold unsupported or unsafe actions.
Human Approval: The user examines the outcome and decides whether to proceed.
The AI agents do not directly calculate financial balances or modify accounts through generated text. Instead, a deterministic simulation engine validates operations and computes outcomes.
Financial simulation engine
Built with Python, FastAPI, Pydantic, and SQLite, the financial engine models banks, businesses, households, credit relationships, payroll, production, and financial shocks.
It records financial transactions in integer cents, checks accounting consistency, and supports reproducible comparisons from saved simulation snapshots.
Selected public economic references, including FDIC and Census data, can inform experimental starting conditions. The financial entities and scenarios remain simulated.
Frontend and deployment
The interface uses React, TypeScript, Vite, and Zustand, with interactive economic visualizations, financial charts, scenario comparison, and agent activity inspection.
The application is packaged using Docker and deployed on Render. NVIDIA model inference runs through Nebius Token Factory.
Challenges we ran into
- Reliable multi-agent orchestration
Coordinating multiple AI agents required managing tool permissions, structured responses, state transitions, and handoffs between independent agent roles.
- Preventing unsupported financial actions
Language models can produce invalid tool calls or recommendations that don't match calculated outcomes. I implemented deterministic checks, independent supervision, human approval requirements, and safeguards against duplicate execution.
- Making financial simulations consistent
Modeling credit, payroll, bank liquidity, business operations, and cascading disruptions required careful accounting and sensitivity testing. I revised assumptions that produced unrealistic baseline behavior while preserving earlier simulation versions for reproducibility.
- Explaining complex outcomes
A technically sophisticated simulation is only useful when people can understand it. I focused on making the consequences of a financial shock, the evidence gathered by agents, and the differences between interventions visible to users.
Accomplishments that we're proud of
Built multi-step NVIDIA Nemotron agent workflows using Nebius Token Factory.
Implemented independent agent supervision and evidence-based approval gates.
Created a financial simulation with balanced accounting checks and reproducible alternate-world comparisons.
Made AI tool interactions, agent collaboration, and review decisions inspectable.
Integrated experimental public economic references with explicit assumptions and data provenance.
Deployed the full-stack application using Docker.
A particularly important result was verifying that when a supervisor held a proposed financial intervention, the application prevented that intervention from being applied.
What we learned
Building an agentic AI system requires more than connecting an LLM to tools. Reliable AI agents need verifiable evidence, carefully controlled actions, deterministic validation, and clear separation between recommendations and execution.
I also learned how challenging it is to build an understandable simulation without overstating what synthetic economic models can predict.
What's next for Econova
I plan to improve Econova through additional economic calibration, real-user usability studies, broader agent reliability evaluation, and more sophisticated financial scenarios.
The long-term goal is to make complex economic interactions easier to explore, understand, and teach.
Try Econova
Live application: https://econova-z7tc.onrender.com
GitHub: https://github.com/shivanginarayan/Econova
Demo: https://youtu.be/CcJSjWJdJWQ
Econova is an experimental educational simulation. Its results depend on documented assumptions and should not be treated as validated forecasts or financial advice.
Built With
- ai-agents
- docker
- fastapi
- financial
- function-calling
- multi-agent-systems
- nebius-token-factory
- nvidia-nemotron
- pydantic
- python
- react
- sqlite
- typescript
- vite
- zustand
Log in or sign up for Devpost to join the conversation.