Resume · Public profile
Salmonbird
Education
Zhejiang University
2022 — 2025Master's · School of Software Technology
Soochow University
2018 — 2022Bachelor's · School of Computer Science and Technology
Profile
AI agent and LLM application engineer. On ByteDance’s Coze team, I worked on core generation pipelines for Coze Space’s PPT Agent and built automated evaluation and self-improvement systems. I focus on turning complex agent capabilities from prototypes into reliable, production-ready systems. As an independent developer, I now build open-source developer tools, AI products, and interactive narrative experiments.
Experience
ByteDance · Coze
Software Engineer
Coze Space PPT Agent
Structured PPT generation
- Owned the end-to-end implementation of the core PPT generation pipeline, spanning content planning, slide generation, local edits, and template reuse, helping move the product from capability validation to delivery at scale.
- Led multiple performance iterations by profiling stage-level latency, identifying bottlenecks, and refactoring critical paths, significantly reducing end-to-end generation time while continuing to improve output quality.
Visual PPT generation
- Designed the architecture from scratch and led its long-term evolution, connecting generation, local edits, template reuse, and PPTX export into a complete product workflow while significantly reducing per-task generation cost.
- Evolved the generation pipeline through multiple architectural stages, including packaging it as reusable Skills and adapting it for commercial delivery, making it a core PPT capability in Coze Space.
- Built an automated evaluation and self-improvement system that replaced manual failure-case analysis with a closed loop spanning multidimensional evaluation, clustering-based root-cause analysis, and iterative validation, continuously improving output usability for core visual styles.
Coze Coding
- Contributed to the “Generate App” agent and owned its local-edit pipeline, replacing full rewrites with incremental generation to substantially reduce output tokens and edit latency.
- Built an evaluation loop spanning builds, execution, screenshot capture, result analysis, and automated repair. Used reproducible execution evidence to drive pipeline improvements and increase success rates for both direct generation and end-to-end delivery.
Open Source
Evoui Browser
Independent Developer / Maintainer
A self-evolving browser execution harness for AI agents
GitHub / Salmonbird/evoui-skill- Designed Evoui Browser to reduce the repeated overhead of having AI agents reread pages, explore paths, locate controls, and validate outcomes from scratch for every web task. Built a self-evolving execution loop spanning Trace → Evolution → Knowledge → Replay → Feedback.
- Built a knowledge layer that selectively distills two complementary forms of knowledge from complete execution traces: Procedure packages verifiable mechanical actions for replay, while SiteMap preserves stable entry points, page mechanics, success and failure signals, and previously disproven paths.
- For later tasks, SiteMap guides agents to the right entry point and away from known dead ends before Procedure replays well-defined mechanical segments. Replay continuously validates applicability, target uniqueness, and successful completion, returning control to the agent whenever the page does not match or the outcome cannot be confirmed. This reduces repeated observation and decision cycles without taking over business-level judgment.
Internships
Alibaba · Taotian Group · Dian Xiaomi (店小蜜)
Backend Engineering Intern
Worked on LLM modules for an intelligent customer service product, focusing on pre-sales scenarios such as product inquiries and introductions.
- Standardized the product-inquiry retrieval pipeline by unifying multiple knowledge sources into a multi-query, dual-channel retrieval and reranking architecture. Added rule-derived selling points such as price and positive reviews, helping improve inquiry conversion rates.
- Built a rapid-response platform for intervening in LLM outputs, covering triggered-response handling, bulk remediation, and dashboards, enabling fast mitigation based on merchant feedback.
- Optimized asynchronous policy templates in the RAG pipeline to automatically surface knowledge-base cards as a fallback when the LLM abstained from answering.
Hundsun Technologies Inc.
University–Industry Program Intern
Designed and implemented a code-migration desktop application supporting a framework migration from WPF to Avalonia.
- Extracted semantic project information and built dependency graphs, applying a three-level cache pattern to resolve cyclic dependencies.
- Implemented three categories of migration rules with Roslyn and used a chain-of-responsibility pattern for extensibility. Supplied an LLM with dependency-graph context and examples retrieved through a dual-channel approach to perform control migration.