nftechie/stonkfly

A full retained fly-connectome simulation with experimental memory and guarded Coinbase AgentKit trading actions.

What it solves

Stonkfly is an experimental simulation of a fly connectome (the map of neural connections in the Central Nervous System) used to operate a crypto trading account. It explores whether a biological neural network structure can be used for decision-making in a financial market context.

How it works

The system converts public Coinbase price data into an RGB chart, which serves as visual input for a simulated brain. This input stimulates 3,335 brightness inputs and 811 color inputs within a graph of 166,700 neurons and 25.6 million connections (MaleCNS v1.0). A fixed readout proposes a trade action (buy, sell, or hold), which is then executed via a Coinbase AgentKit ActionProvider.

Learning is handled through engineered reinforcement signals: positive profit and loss (P&L) stimulates dopamine cells, while negative P&L stimulates aversive dopamine cells, triggering a memory rule that modifies connections between Kenyon cells (KC) and Mushroom Body Output Neurons (MBON).

Who it’s for

Researchers or developers interested in biological neural network simulations, connectomics, and experimental AI trading bots.

Highlights

  • Uses a real fly connectome graph (MaleCNS v1.0) with over 166,000 neurons.
  • Integrates with Coinbase Advanced for both paper trading and live spot orders.
  • Implements biological reinforcement learning via simulated dopamine cell stimulation.
  • Includes a safety-first setup with default order limits and drawdown stops.

Related

  • Project
    hybridgroup/yzma

    A Go library that integrates llama.cpp for fully local LLM and VLM inference with hardware acceleration, requiring no CGo or external servers.

  • Project
    aovestdipaperino/tokensave

    TokenSave is a local MCP server that builds a semantic knowledge graph of a codebase, letting AI coding agents query symbols, callers, and impact instead of scanning files with grep and read, cutting token usage and tool calls.

  • Project
    datalorax/equatiomatic

    An R package that automatically converts fitted statistical models into LaTeX equations, reducing the manual effort of writing mathematical formulas for research papers.

  • Project
    warpfront/hipfire

    A fast, AMD-native LLM and image generation inference engine written in Rust and HIP that removes Python from the hot path to optimize performance on RDNA and CDNA GPUs.