Inspiration

Reasonance was inspired by Saga Anderson’s Mind Place investigation board in Alan Wake 2. Watching clues, evidence, and theories get arranged across a physical board over time revealed something important: real reasoning is not a single question and answer — it is iterative, visual, and accumulative.

Most AI tools today still behave like chat interfaces. But humans don’t reason in chats — we reason by forming hypotheses, testing them against evidence, revisiting earlier ideas, and refining our understanding across time. Reasonance was born from the idea of turning that detective board into a living, multi-turn reasoning system powered by Gemini.

What it does

Reasonance is a multi-turn abductive reasoning engine where Gemini conducts investigations across structured turns instead of single prompts.

Users upload source material (articles, case files, paradoxes, research papers), and the system:

  • Generates initial hypotheses and evidence (Turn 1)
  • Allows users to select a hypothesis and link evidence
  • Uses Gemini to generate “What If” explorations, new conclusions, and cross-references
  • Builds a persistent reasoning graph across turns
  • Synthesizes the entire investigation into a structured case file

Instead of chat history, Reasonance creates a detective board of evolving thought.

How we built it

Reasonance is a full-stack application:

  • Spring Boot + PostgreSQL store investigations as graph structures (nodes and edges)
  • Angular + TailwindCSS render a turn-based investigation board inspired by Alan Wake 2
  • Gemini 3 acts as the reasoning engine
  • Ollama (gemma3n) was used for local experimentation and prompt iteration

The core architecture is a turn loop:

  1. Gemini generates hypotheses and evidence from the source
  2. User selects a hypothesis and links evidence
  3. Backend reconstructs the full reasoning graph and sends it to Gemini
  4. Gemini produces new hypotheses that explicitly reference earlier turns
  5. This repeats across multiple turns (limited to 5 for MVP) before a final synthesis

This transforms Gemini from a responder into an investigator.

Challenges we ran into

  • Getting Gemini to maintain continuity across turns without drifting
  • Designing strict JSON schemas so reasoning could be parsed and stored as a graph
  • Preventing the system from becoming just a “prompt wrapper”
  • Building a UI that feels like a detective board, not a dashboard
  • Forcing explicit cross-referencing between turns
  • Balancing user-guided exploration with AI-generated reasoning
  • Tried building similar investigative board as in the game, but it was not suitable for the reasoning engine developed

Accomplishments that we're proud of

  • Turning Gemini into a multi-turn reasoning partner, not a chatbot
  • Building a visual reasoning board that changes how users think
  • Creating a persistent hypothesis graph across investigations
  • Achieving structured abductive reasoning with explicit lineage
  • Using various agents to develop the application. Not just using claude code or codex. Using bunch of the agents orchestrating at a time to build this.

What we learned

  • LLMs are powerful at hypothesis generation but require structure for continuity
  • True reasoning needs state, memory, and architecture, not just prompts
  • Visualizing reasoning changes how users interact with AI
  • Gemini’s long context window enables reasoning lineage across turns
  • Prompt engineering alone is not enough — reasoning systems need design

What's next for Reasonance

  • Support for PDFs, audio, and video as investigation inputs
  • Automatic visualization of the reasoning graph
  • Collaborative investigations across multiple users
  • Exportable investigation reports
  • Deeper integration with Gemini’s multimodal and long-context capabilities
Share this project:

Updates