Jun Yu Tan is a founding engineer at Tusk who builds coding agents, automated testing systems, and security tools that make AI-assisted software development more dependable. He advocates AI that strengthens human judgment, exposes uncertainty, and verifies its conclusions before asking developers to act.
Tan graduated from the National University of Singapore with a degree in data science and analytics. He worked on data and machine-learning projects at Autodesk and Micron, where he researched acoustic anomaly detection for manufacturing equipment, before becoming a machine-learning engineer at Saleswhale, a conversational-AI company acquired by 6sense. He began pursuing a computer-science master’s degree at Georgia Tech in 2023, joined Tusk in July 2024, and started researching social agent-mediated interactions at Georgia Tech’s Design Intelligence Lab in January 2025.
Shortly after joining Tusk, Tan coauthored the release of its Recursive Agent, which separates validated assumptions from risky unknowns and sends specialized subagents to investigate uncertainties concurrently. His subsequent work addresses what happens when increasingly capable agents encounter real codebases, production traffic, and limited human attention.
- Systematic pessimism. Tan uses AI to inspect changed code, trace downstream effects, identify edge cases, and generate tests that actually run against the relevant environment. His approach to software failure turns speculative warnings into reproducible bugs and regression tests.
- Human-centered AI augmentation. Tan argues that coding assistants should surface blind spots, explain their assumptions, accept correction, and preserve human responsibility for consequential decisions. His AI Engineer World’s Fair presentation also advances novelty-criticality prioritization: unfamiliar, serious risks deserve interruption; routine reminders and cosmetic suggestions should consume less attention. He measures successful assistance partly by whether users become more capable over time.
- Constraint-first testing with Tusk Drift. While developing production-traffic replay tools, Tan designed request matching around HTTP semantics, execution stage, and whether a recorded response has already been consumed. His constraint-based mock matching eliminates invalid candidates before using similarity scores to resolve ties, improving deterministic replay and regression detection.
- Fence. Tan built Fence, an open-source, container-free sandbox that restricts network access, filesystem operations, and commands. Its default-deny network policy establishes enforceable boundaries for coding agents and other semi-trusted processes.
Tan has also examined efficient byte-pair tokenization and modern transformer architectures, extending his interest in reliable AI from developer-facing safeguards to underlying model infrastructure.
