Planned coursework: EECS 280 (Programming & Data Structures), EECS 203 (Discrete Mathematics)

Relevant coursework: Linear Algebra, Calculus I–III, Probability & Mathematical Statistics, Differential Equations, Introduction to Proofs, Discrete Structures, Computer Programming

• Trained LoRA adapters on a DiT-based text-to-image model (Anima, a Cosmos-Predict2 derivative) using diffusion-pipe; systematically compared output quality across learning rate, LoRA rank, and captioning strategy variations.
• Built an automated dataset annotation pipeline using vision-language models (Qwen-VL) to generate structured captions for illustration-domain training data, replacing manual tagging workflows.
• Developed a custom ComfyUI node for XY-grid parameter sweeps, enabling controlled visual evaluation of generation quality across hyperparameter axes.
• Analyzed model architecture: identified that character and style knowledge concentrates in a small LLM adapter on the text encoder rather than in the diffusion backbone, consistent with but independent of developer findings.

• Architected and built a self-hosted AI platform with modular design: multi-provider LLM backend, tool-use subsystem, and cross-platform adapters (Discord, Telegram, email via IMAP).
• Designed Symphony, a DAG-based visual automation engine enabling users to compose multi-step LLM-powered workflows without writing code.
• Implemented a retrieval-augmented memory system using dual-layer entity graphs with Personalized PageRank scoring for context-aware recall.
• Built a cross-platform computer-use module for LLM-driven desktop interaction via native OS APIs (C# on Windows, Swift on macOS).

• Constructed a 2,084-turn character dialogue dataset by reverse-engineering the MAGES visual novel engine, extracting and processing 84,000+ lines of game scripts.
• Conducted supervised fine-tuning on Qwen3-8B-Base; evaluated character voice consistency and multilingual coherence across training configurations.
• Benchmarked dense open-weight base models on multilingual tokenization efficiency to inform architecture selection for bilingual dialogue generation.