Xiang Yashan
Software Engineering · Big Data
Focused on LLM agents, RAG, and enterprise knowledge-base engineering. At Rayneo, worked on voice-assistant intent recognition and long-term memory research whose capabilities shipped in the mass-produced RayNeo iO AI glasses; shipped a production-grade CAP application at SAP and drove the platformization of multimodal document intelligence at BSH. Fluent in Python and PyTorch; strong at turning business pain points into deployable AI solutions.
Timeline
Experience
- AI voice assistant & intent recognition: owned intent recognition and temporal parsing for the voice assistant, covering weather, lunar calendar, reminders, countdowns, device control, and more, plus a self-developed temporal-parsing engine (Gregorian / lunisolar / time-range, with a full unit-test suite). Designed a device-cloud hybrid architecture that dynamically routes between an on-device BERT intent model and cloud LLMs via long/short-horizon model routing to balance latency and comprehension, packaging the capabilities as low-latency, Dockerized distributed microservices; also built a Multi-Agent Self-Evolve (MASE) data-synthesis framework that closes the loop on on-device data collection, training, and iteration to steadily improve sample diversity and intent coverage, with core capabilities exposed as MCP tools for agent invocation. These capabilities shipped in RayNeo iO (Aug 2026), the industry's first "human-augmentation AI glasses".
- Long-term AI memory research: contributed to "life-memory" research for always-on smart-glasses scenarios, building LifeDialBench — a long-term memory benchmark with a real first-person EgoMem subset and a year-long simulated LifeMem subset — proposing a top-down hierarchical life-simulation framework, generating event-tree-based QA across multiple time scales, and designing an Online Evaluation streaming protocol that mirrors real all-day wear; under a unified evaluation, systematically compared the retrieval accuracy and temporal reasoning of mainstream memory frameworks (RAG / A-Mem / Mem0 / MemOS). One of the core efforts of RayNeo's AI Memory team research, it underpins RayNeo iO's "all-day memory engine".
- Intelligent knowledge base & ticket AIOps: for smart-glasses usage scenarios, designed and implemented a customer-service knowledge-base system fusing LLMs with Milvus (2,700+ knowledge entries) with semantic retrieval, version management, and RAG-based Q&A, powering AI-driven knowledge lookup for user support and community ops; and built an intelligent triage pipeline for real fault tickets (1,700+ tickets) where AI classifies and anonymizes issues, then auto-routes them across ticketing systems to the corresponding Feishu owner tables with owner notifications — turning manually-judged dispatch into an automated closed loop that markedly shortens the response chain.
- Shipped a production-grade application: designed and implemented the development and deployment of a cloud-native CAP application on the SAP BTP platform, automating the enterprise data-transformation and printing workflow. Responsible for service-interface design and deployment monitoring, running stably in production with a 30%+ improvement in process efficiency.
- Accumulated enterprise-grade development experience: gained hands-on mastery of Node.js application and LLM integration development, delving into business processes and designing efficiency-boosting solutions leveraging AI tooling. The project spanned the complete pipeline from requirements analysis to deployment monitoring, delivering solutions for enterprise AI digital transformation.
- Designed and built multimodal document intelligence processing: for diverse document types such as recipe data and interaction reports, designed and developed an LLM-Agent-based automated processing pipeline delivering information extraction, tag generation, content summarization, and more, achieving an 80%+ recall rate.
- Drove the platformization of AI algorithms: developed the StreamPilot production solution on Langflow, designing multiple reusable processing components that enable drag-and-drop algorithm reuse across scenarios, turning technical solutions from code into configurable platform services.
Education
State Key Lab for Novel Software Technology · NJU — Terrain3DSketch
Research on accelerating 3D authoring and modeling on AR platforms. Responsible for the deep-learning generative algorithms and terrain data processing, and developed the Unity/HoloLens-based Terrain3DSketch system, which lets users sketch mid-air and automatically generates controllable, personalized terrain models.
Jiangsu Provincial Key Lab of Wireless Sensor Networks · NJUPT — RF Non-Contact Vital-Sign Monitoring
As principal investigator, completed a provincial-level Undergraduate Innovation project: an RF-based non-contact intelligent vital-sign monitoring system, which earned an Outstanding completion evaluation. My undergraduate thesis, built on this work, was named a Jiangsu Provincial Outstanding Thesis. Responsible for the deep-learning core algorithms and signal-data acquisition and processing, and built a complete non-contact heartbeat-detection system with the Streamlit visualization frontend, reaching 93.05% heart-rate computation accuracy at low latency.
Research
Awards
Awarded as team leader.
Awarded by the Far East Holding Group.
ATA AutoCAD Application (Mechanical) Intermediate certification; Jiangsu Computer Rank Exam Level-2 C Programming (Distinction); National Computer Rank Exam Level-2 C Programming certificate.