Hi, I'm AJ, an aspiring Research Engineer. I generally enjoy mechanisms in any concepts and got into computers. So I decided, let's go explore the mechanisms in computers, but also build new things with the learned mechanisms, but I learned computers can think so let's do a cross-over
To sum up, I'm generally curious on how things work and tend to chase novelty. But I'm mainly under the AI field.
- Applied LLM Engineering Projects
- LLM Infrastructure
- AI Companion Memory
- AI Companion Emotion
- AI Session Emotion Classifier
- C++ Software Development
- Lighweight Tag-Based Text Files Database
Project links are below
Working with post-trained LLM APIs and building softwares around them.
Systematic and architectural vision, understands both micro and macro-level.
• MemoryGate WIP: A lightweight memory layer for roleplay chat agents
• AetherRouter WIP: Async, multi-provider LLM routing and fallback engine
• Syneuro WIP: A lightweight Quantum-inspired Emotion Layer for AI Chatbots
Building systems from first principles in C++: database engines today, deep learning and ML systems next.
Interested in what runs underneath: memory, storage, concurrency, and the compute that powers models.
• KojinDB WIP: A lightweight personal database engine (個人, kojin = "personal") inspired by PostgreSQL.
Co-authored two papers, contributed to the paper structure, research questions, and running experiments.
Conceptual co-author, with experience in synthesizing literature into a coherent research question.
• First-Token Broadcasters: Mechanistic Origins of Language Identity and Distributed Robustness in Transformers: Paper on the mechanistic origins of language identity and distributed robustness in Transformers.
• FTB Graph: Determining and Validating First-token Broadcasters and Language-Identity Head Circuits in Multilingual Language Models: Paper on determining and validating first-token broadcasters and language-identity head circuits in multilingual LLMs.
Worked with an experiment codebase for imitation learning, contributed to the infrastructure side.
Infrastructure-oriented, understanding data transformation in experiment workflows.
• jetspace-imitation-learning:
- Dataset Loading Optimization - Optimized dataset loading by implementing options-based selective NPZ decompression and adding fail-first guard clauses to skip invalid episodes early.
- Data Pipeline Refactor - Refactored data pipeline architecture by decoupling Episode into a standalone dataclass, upgrading EpisodeBuffer storage mechanics, and updating the demo collection script for compatibility.


