Naoki Shibuya

ShinkaEvolve-FE is a proof-of-concept financial risk-analysis system built on top of SakanaAI’s ShinkaEvolve. It demonstrates how Shinka can evolve a reusable prompt-building function that automatically generates consistent, scenario-specific hedge-weakness analysis using LLMs. The system does not replace pricing or optimization. Instead, it automates the natural-language analysis that normally follows quantitative stress results.

In this POC, a simple Greek-based optimization model produces worst-case multi-factor crisis shocks under realistic per-factor sigma limits. Shinka consumes these quantitative outputs and produces structured prompts that guide an LLM to explain why the stress scenario is harmful, how hedges behave under crisis correlations, and which risk factors dominate the loss.

The quality of each analysis is then evaluated by a separate reviewer, LLM, who scores the output and provides feedback. These scores serve as the fitness signal for Shinka’s evolution process, allowing the prompt-building function to improve over generations.

This illustrates a practical separation of responsibilities: quantitative models compute the scenario, and Shinka automates and iteratively improves the explanatory analysis layer.

Multi-agent chat system featuring LLM-based intent routing, response evaluation with agent forwarding, native tool calling, MCP integration (finance, weather, web search), topic-aware RAG with embeddings, real-time streaming (SSE), and a multi-turn chat UI with edit-and-restart. Agents are fully configurable via YAML. Supports Ollama (local), Gemini, Claude, and ChatGPT through a unified provider interface.

I've written a series of articles on the transformer architecture and language models on Medium. This repository contains an implementation of the Transformer architecture presented in the paper Attention Is All You Need by Ashish Vaswani, et. al. My goal is to write an implementation that is easy to understand and dig into nitty-gritty details where the devil is.

The objective of this project is to provide a Python-based framework for optimizing wire-current schedules on atom chips. The framework is designed to facilitate the transport of Bose–Einstein condensates (BECs) while maintaining the desired trap geometry and ensuring adiabaticity. It leverages gradient-based optimization techniques to compute optimal current schedules efficiently.

Built and trained a convolutional network for end-to-end driving in a simulator using Tensorflow and Keras. I implemented the driving (steering and throttle) model based on the Nvidia paper on deep learning and self-driving cars and using Udacity's car simulation environment for training and testing.

Link to the Nvidia paper used for this project:
https://devblogs.nvidia.com/parallelforall/deep-learning-self-driving-cars/

This blog article explains how upsampling with transposed convolution works for those who may be confused about the logic. This is one of the posts I wrote for Medium, and it was the most popular one.

I've written different deep learning examples to accompany some of my Medium blog articles (http://naokishibuya.medium.com). I covered the topics such as style transfer, autoencoder, generative adversarial network (GAN), and transposed convolution.