Systems

Jouzu — LLM-Driven Character Learning Platform (Founder)

Website: jouzu.ai

A production Japanese learning application centered on stylized conversational agents.

• Designed persona-conditioned dialogue agents for language learning
• Integrated speech input/output for conversational practice
• Built and deployed backend APIs and learning workflows
• Led cross-functional team (engineering, art, voice)

Research focus: engagement, adaptive dialogue, long-term interaction.

VoxRAG — Transcription-Free Spoken Question Answering

GitHub: zackrack/voxrag
Publication: MAGMaR @ ACL 2025

Explores retrieval-augmented generation directly over speech representations rather than relying on full ASR transcripts.

• Built multimodal retrieval + generation pipeline
• Designed speech-grounded evaluation framework
• Investigated robustness to ASR noise

AdaptLingo — Adaptive Speech-to-Speech Dialogue Agent

GitHub: zackrack/adaptlingo

End-to-end conversational system that adapts responses to user English proficiency.

• Integrated ASR, prosody analysis, constrained decoding
• Implemented speech-to-speech conversational loop
• Focused on real-time adaptive difficulty control

RAG Product Assistant — Infineon Technologies

(Code proprietary)

Enterprise retrieval system for technical documentation.

• Designed document ingestion + indexing pipelines
• Integrated RAG architecture into internal workflow
• Reduced product lookup time from minutes to seconds

Re:Member — Emotional Question Generation System

Publication: HCI+NLP @ EMNLP 2025

Dialogue system that generates emotionally-aware questions from personal memories.

• Designed LLM-based generation framework
• Investigated affect-aware prompting
• Evaluated question diversity and emotional alignment

Expressive Japanese TTS Evaluation Pipeline

Publication: IEEE UEMCON 2025

Comparative evaluation framework for expressive multilingual TTS models.

• Benchmarked VITS and Style-BERT-VITS2
• Designed human evaluation study
• Built analysis pipeline for speech quality metrics

Jouzu Engagement Analysis System

Built data collection and analysis pipeline for studying learner-agent interaction outcomes.

• Logged conversational patterns
• Analyzed engagement metrics
• Connected usage patterns to learning outcomes

Speech-Based CALL Prototypes (Columbia / University of Tokyo)

Research prototypes exploring:

• Speech-integrated dialogue agents
• Multimodal interaction in educational settings
• Adaptive conversational scaffolding