动漫技术宅

十年独立站,分享生活和思考。

用动漫诠释生活,用技术改变未来。

Curriculum Vitae

Algorithm Engineer · CV & LLM

End-to-end experience across CV/LLM training, instruction tuning, RLHF, reward modeling, and production deployment for enterprise AI systems and agents.

Computer Vision LLM / Agents RLHF · Evaluation · Research-to-Production

About

Algorithm engineer focused on computer vision and large language models. Experienced in detection, segmentation, classification, and tracking, with hands-on work from model training and instruction tuning to RLHF, reward modeling, and production deployment. Led AI agent development end-to-end—policy design, reinforcement learning optimization, and enterprise rollouts. Holder of multiple patents and papers, with strengths in cross-task modeling, complex system engineering, and model capability evaluation.

Education

  • Master of Computer Science
    — · National University in Taiwan (QS 200+)

    Led multiple computer vision research projects and transferred outcomes into real-world deployments.

  • Bachelor of Computer Science
    — · National University in Taiwan (QS 200+)

    Focused on computer vision research and applied systems.

Work Experience

  • Algorithm Engineer · AI Lab, Shanghai (Company name withheld)
    2022 - Present

    Leading CV, LLM, agent, and reinforcement learning systems from core algorithms and training pipelines to evaluation and production deployment. Delivered 10+ vision systems supporting 60+ internal and external scenarios; improved zero-shot vision metrics by 15–40% and boosted inference speed by 30%+. Raised agent workflow success from 40% to 85%+, built a multimodal evaluation suite that tripled regression efficiency, and performed safety tuning across 20+ scenarios and 1,000+ risk patterns, reducing false alarms by 25%.

Projects

Multifunction Visual Sandbox
Zero-shot Detection · SAM · Multi-task Vision · Streaming

Built a unified sandbox integrating 30+ vision models (detection/segmentation/classification) with a single inference interface; supported RTSP/RTMP streams with real-time visualization and pluggable benchmarking; reduced validation cycle from 3 days to hours and increased deployment efficiency by 200%+.

Agent Workflow Auto-Orchestration
LLM · Agents · Function Calling · RLHF

Designed task decomposition, planning, and tool-calling logic with self-correction; built feedback-driven fine-tuning (SFT + PPO/RLOO) for closed-loop optimization; improved success rate from 40% to 85%+, cut average steps by 25%, and reduced workflow generation time by 60%.

Zero-data Auto Training System
Synthetic Data · Prompt Training · Bootstrapping

Created a fully automated pipeline from task definition to data generation and model training with no manual labeling or pretrained models; reduced labeling cost by 90%+ and achieved 10–25% gains over common zero-shot baselines on target tasks.

Visual Agent Benchmark Suite
Open-ended QA · VLM Reasoning · Hallucination

Built a multimodal evaluation framework for agent reasoning, cross-image analysis, and structured output consistency; added uncertainty scoring and reward-model feedback, tripling regression speed and improving error localization by 40%+.

Publications & Patents

Papers (Confidential)
Research to industry transfer

Authored multiple papers with results transferred to production; details are not publicly disclosed due to privacy.

Patents (Confidential)
Multiple filings · CV / LLM / Agents

Multiple patent filings related to cross-task modeling, system architecture, and evaluation; details withheld for confidentiality.

Contact

  • Email: h@dmjsz.com
  • Web: dmjsz.com