Tong Pei

Led the end-to-end development of an Agent Memory SaaS platform, including public deployment, API encapsulation and remote dynamic memory read/write access for external agent nodes. Designed and implemented LLM-driven agent systems with LangChain, enabling natural-language control and protocol-level integration for Xiaomi smart-home devices and a broad range of cameras. Built task-specific prompt orchestration and workflow pipelines using an in-house Harness framework, reducing hallucination rate by 20% in defined scenarios and improving contextual memory consistency. Developed a memory-system plugin for the OpenClaw framework, enabling automatic account registration and autonomous memory data read/write; improved long-context complex-task memory performance by 40%. Contributed to an AIoT system for smart furniture and developed a patent-pending rail-mounted operating structure for electronic products.

Developed and optimized object detection and facial recognition models using YOLOv3 and U-Net, improving recognition accuracy by 53% compared with the baseline model. Contributed to the full ML development lifecycle, including data collection, cleaning, augmentation, labeling, model training and hyperparameter tuning.