Selected work   /   Beyond the agent loop   /   All repositories

Gwanyu Park

Machine Learning · Language Models · Agent Systems

Research-minded. Systems-driven.

I work at the intersection of machine learning research, language-model engineering, and entrepreneurship, with experience across professional services, legal technology, and biotechnology.

My foundation is an Advanced Computing (Honours) degree in Computer Science from the University of Sydney. My experience includes ML research at Ernst & Young in Sydney, language-model development at ChangLaw Group in the United States, and applied ML work at a biotechnology research center.

Beyond these roles, I built and ran my own business focused on machine learning and big data. That work led to a granted Korean patent in big-data search technology: KR102195556B1 (2020).

My technical interests span machine learning, deep learning, reinforcement learning, and language-model quantization. I also build agent harnesses as personal projects, exploring the systems around language models: orchestration, memory, tool use, and evaluation.

I’m drawn to the work between an interesting idea and a system you can actually inspect, test, and improve.

Selected work

Three parts of the same question: how do we make agent work easier to inspect?

Weasley Subagents ↗ — Coordinate focused agents without losing the parent session. A Pi extension with parallel reviews, explicit handoffs, background runs, and bounded execution.
_{TypeScript · Pi · Worktree isolation}

Weasley DeepMind ↗ — Give agents memory with structure. A self-hosted MCP server for typed fragments, linked evidence, and retrieval within a token budget.
_{JavaScript · PostgreSQL / pgvector · MCP}

Weasley AutoResearch ↗ — Make experimentation leave evidence. A guarded Pi loop that runs benchmarks, records results, keeps improvements, and reverts regressions.
_{TypeScript · Pi · Benchmark harnesses}

Beyond the agent loop

• Weasley Assistant — A local-first Python runtime with explicit tool calls, layered memory, traces, and deterministic evaluation.
• Weasley Podex — A multi-agent workspace and orchestration toolkit with a browser dashboard.
• potato box — Source material into a linked Markdown knowledge base, with provenance and review queues.
• Potaty — Editable ASCII diagrams grounded in prompts, transcripts, and repository evidence.
• Weasley Frame — Live-web research into sourced datasets you can inspect, refresh, and export.

Questions behind the code

Control — How can agents delegate work while keeping scope, state, and handoffs explicit?

Continuity — What should an agent remember, and what evidence should travel with it?

Feedback — How can we measure an improvement and retain the record of what changed?

---

_{Open the repositories. Follow the evidence. Build on what is there.}