NashNode

Inspiration

NashNode started when I was deciding between a fixed or variable mortgage in Canada.

The decision depended on several connected factors:

• Inflation
• Interest rates
• Oil prices
• Geopolitical conflicts
• How governments and markets might respond

Game theory was a good way to model these interactions, so I wanted to combine it with AI and visualize the possible decisions and outcomes.

What it does

NashNode converts a real-world situation into a multi-round strategic game.

It can:

• Identify the players, goals, preferences, constraints, and possible moves
• Let users define what “winning” means for them
• Predict the user’s and opponent’s future actions
• Generate an interactive decision tree
• Calculate the probability and expected value of different outcomes
• Highlight the user’s win path, best-value path, and most likely path
• Recalculate the entire tree when the user changes a move
• Update predictions when real-world events happen

How we built it

We first defined:

• The product requirements
• The game-theory and simulation logic
• The Firebase data model
• The decision-tree UX
• The interaction between confirmed history and predicted actions

We documented the design and used Codex in an iterative, goal-based development loop to build and refine the application.

Challenges we ran into

• The AI sometimes generated generic or meaningless actions
• Different branches were too similar
• The initial setup required too much information
• Large decision trees became difficult to understand
• Opponent actions were not clearly separated from user decisions
• Probability estimates needed to be useful without appearing falsely precise

We improved this by:

• Asking only high-value follow-up questions
• Generating strategically different moves
• Limiting visible branches
• Separating confirmed history, the current decision, and predicted future actions
• Showing the assumptions and confidence behind predictions

Accomplishments that we're proud of

• Turning an unstructured problem into a structured game model
• Supporting user-defined win conditions
• Predicting multi-round user and opponent actions
• Calculating terminal outcome probabilities
• Recommending a conditional strategy tree instead of only one next move
• Allowing users to change a move and immediately see how all outcomes change
• Allowing real-world progress to advance and update the simulation

What we learned

• How classic game-theory models can support real-world decisions
• How to combine game theory, behavioral assumptions, and probability updates
• How important structured context is for AI output quality
• How to simplify complex decision trees through better UI and UX
• Why users should define their own success criteria instead of using a universal definition of winning

What's next for NashNode

• Build a remote MCP version
• Publish NashNode as a GPT app
• Improve probability calibration using predicted and actual outcomes
• Add reusable templates for negotiations, financial decisions, career choices, and business strategy
• Integrate documents, email, and external data
• Let Codex and other AI agents use NashNode as a persistent strategic reasoning engine

Built With

• chatgpt
• codex
• firebase
• llm
• typescript