Hikaru Kondou

About Me

I'm Hikaru Kondou. I have a strong preference for Rust programming language and am particularly interested in deep learning acceleration and high-performance computing (SIMD, CUDA).

Work Experience

Parakeet, Inc.

Employment Period September 2024 – Present
Business Software development, voice conversion technology
Project: paravo - Real-time Voice Conversion App
Role Lead Developer & Project Manager
Technologies Rust (Tauri), TypeScript, React
Overview Leading the development of paravo, a real-time voice conversion application that has reached 60,000+ downloads (as of March 2025).
Responsibilities
  • Project Leadership: Served as the primary responsible person for the entire development process.
  • Software Architecture: Designed and implemented the core architecture, ensuring performance optimization for real-time voice processing.
  • UI/UX Design: Led the design process to create an intuitive interface for voice conversion functionality.
  • Cross-platform Development: Utilized Tauri framework to deliver a consistent experience across different operating systems.
Achievements
  • Grew user base to 60,000+ downloads
  • Successfully implemented real-time voice conversion with minimal latency
  • Created a cross-platform application with native performance using Rust and Tauri
  • Achieved industry-fastest conversion speed with only 0.08sec latency

ax Inc.

Employment Period October 2022 – September 2024
Business Deep learning consulting and application development
Founded 2018
Capital 352 million JPY
Employees 50
Project 1: Game Voice Verification Tool (April 2024 – July 2024)
Role Software Engineer / Machine Learning Engineer
Technologies Python, Whisper, Torch, Hugging Face, ESPnet
Overview Developed a voice verification tool for a game production company to check if recorded voice lines match the original script. This project addressed challenging short utterances like "Uwa—!" or "Yaa—!", which are difficult to recognize using conventional speech recognition systems.
Achievements
  • Improved Speech Recognition Accuracy: Fine-tuned a Whisper-based model using katakana-only datasets. Recognition accuracy improved to ~96% for typical lines, and short exclamations like "Uwa—!" reached ~50%.
  • Handling Inconsistent Data: Proposed a workflow to flag uncertain transcriptions and created an auto-standardization tool, reducing manual correction time.
  • Training Scripts: Delivered scripts that allow fine-tuning with proprietary data, enabling clients to re-train the model with their own datasets.
Team Structure 1 Project Manager, 1 Sales Representative, and 1 Engineer (yourself)
Project 2: DX Application Development (October 2023 – Present)
Role Software Engineer
Technologies Flutter 3.19.6, C++17
Overview Contributed to the development of a DX application using the in-house ONNX runtime called "ailia." Major achievements included creating an OCR library from scratch and implementing a meeting minutes feature.
Achievements
  • OCR Library Development: Built an OCR library from the ground up, including image preprocessing, without relying on libraries like OpenCV (to avoid large binary sizes). Achieved efficient and high-quality image processing while meeting strict size constraints.
  • Meeting Minutes Feature: Utilized Flutter's isolate functionality for parallel processing and implemented responsive state management with ChangeNotifier.
Team & Role A total of 15 engineers on the project; you served as both an implementation engineer and a code reviewer.
Project 3: ailia Development (September 2022 – November 2022)
Role Software Engineer
Technologies C++17
Overview Implemented previously unsupported layers for the in-house ONNX runtime "ailia."
Project 4: Rust Binding for ailia (November 2022 – December 2022)
Role Software Engineer
Technologies Rust
Overview Created Rust bindings for "ailia," the company's proprietary ONNX runtime.

Tosho Inc.

Employment Period March 2024 – Present
Business Deep learning consulting and application development
Founded 1971
Capital 1,000 million JPY
Employees 622
Project: Medication Verification Tool (March 2024 – Present)
Position Contract-based / Machine Learning Engineer
Technologies Python, PyTorch
Overview Developing a software tool to verify if medications match prescriptions before packaging.
Achievements
  • Medication Recognition System: Used instance segmentation to handle a large and growing variety of medications. Achieved over 95% accuracy, significantly improving workflow efficiency.
  • Code Management Improvements: Introduced GitHub for version control, enhancing collaboration and visibility.
Team Structure 1 Project Manager, 1 Machine Learning Engineer (yourself)
Challenges & Solutions Challenge: The huge variety of medications and overlapping pills made direct classification impractical.

Solution: Implemented instance segmentation for shape and color extraction, then matched to a database. Future plans include OCR for pill text and overlap detection to further boost accuracy.

Technologies

Tauri Hono CUDA

Project

github.com/bokutotu/zenu deep learning framework.

github.com/bokutotu/copilot-client.rs Copilot client for Rust.

Education

Undergraduate: Tohoku University, School of Engineering, Department of Electrical, Information and Physics Engineering (2016/04-2020/03)

Graduate: Tohoku University, Graduate School of Information Sciences (2020/04-2022/03)