○ Seedance capability development: Advancing caption-data strategy optimization and building an automated workflow through implementation and launch. Further details are omitted under the applicable confidentiality agreement.

○ DM model EQ launch: Contributed to SFT for Qwen3.5-397B, raising its EQ score from 50.07 to 64.37. Combined with lead-capture frequency optimization, the launch delivered a statistically significant +4.54% lift in CVR2.

○ DM model v2.0 iteration: At the foundation-model layer, identified Qwen3.5-397B's high cost, lack of audio/video capabilities, and latency constraints; helped restart the foundation-model and workflow discussion and define evaluation dimensions covering DM capability, cost, latency, multilingual support, and multimodality. At the evaluation layer, diagnosed overlapping v1.0 EQ rubrics and an LLM judge that was not aligned with human review; contributed to rebuilding Golden Data and the rubric system from scratch, stratifying SFT training data and benchmark distributions by user intent × vertical, and introducing an iterative human-model alignment workflow. The first Golden Data version has been completed.

○ Agent flow optimization: Contributed to launching a dynamic knowledge base and memory mechanism for merchants' personalized bot-reply styles. The system summarizes business knowledge and messaging preferences on a T+1 basis from advertiser webpages and real conversations, addressing the stateless nature of LLMs. Led SBS human evaluation, which showed materially stronger cross-session coherence and personalization than the baseline, and advanced the solution into an A/B test.

○ Machine-evaluation interception: Contributed to an S0-S3 bad-case severity framework and rolled out full machine evaluation for S0/S1 cases, reducing latency from 16 seconds to 5 seconds while meeting safety requirements.

○ User journey optimization: Identified lead loss outside business hours and launched a configurable night-shift mode to take over high-potential nighttime leads and recover otherwise lost customers.

○ Short drama generation platform: Led the iteration of Nebula AI short drama generation workspace, implemented Master-Sub-Worker Agent architecture and HITL mechanism, connected script structuring, asset generation, storyboard and video generation workflows, unified management of Workflow Snapshot and Provider Job status, achieved partial repainting and breakpoint resume, reduced computing power loss, and supported product Dogfooding.

○ Short drama workflow evaluation: Coordinated the effect evaluation closed loop of AI short drama workflow, designed and carried out SBS evaluation at both Clip and Episode granularities, continuously guiding the tuning of Agent workflows.

○ Digital-human model data preparation: Contributed deeply to the business definition of a 1.6B talking-head model. Built a live-stream micro-action dataset from scratch, systematically defining proprietary head, micro-expression, and hand-action standards together with prompt-mapping logic; called Seedance 2.0 and Kling v3 at scale to batch-generate training videos.

○ Bernini video-editing model evaluation: Advanced SBS evaluation for source-video + prompt video-redrawing tasks. Built an evaluation set of nearly 1,300 cases, benchmarked Bernini closely against Kling V3 Omni, Wan 2.7, Grok Imagine, and other models, calculated human-preference win rates through blind labeling, provided a quantitative view of model capability, and selected showcase examples.

○ Short drama market Agent: For overseas markets, transformed multi-source unstructured data into standardized short drama genre clustering weekly reports, and accumulated visual recognition, weight calculation, genre labeling and quality review mechanisms.

○ Core achievement: Delivered the AI recording bracelet “Memora” from 0 to 1 and passed real testing by 28 executives (cumulative recording 7634h, daily average >7.5h), building Xiaomi AI differentiation barriers with “on-device capability + Human-Car-Home ecosystem Agent”; and advanced the breakdown of commercial model and KOL pilot plans.

○ Multimodal perception and end-to-end understanding: Used Gemini 2.5 Pro (cloud) / MiMo-Audio (on-device) to conduct end-to-end understanding of day-level audio, outputting time-location-event structured scripts; established event segmentation benchmark, with average Precision 0.67 / Recall0.85 / Temporal IoU 0.63.

○ Memory construction and proactive intelligence: Designed and implemented hierarchical storage architecture, stored episodic memory based on file architecture, and extracted SVO triples through Graphiti to build graph-based semantic memory. Based on daily event flow, designed Agent-driven plugin distribution strategy to support personalized recall and push, and implemented sliding card interaction, with favorite rate increased by 25%.

○ Hardware engineering and data link optimization: Deeply participated in the definition of Wi-Fi transmission protocol based on ESP32-S3, and tackled technical problems around binary stream upload, sequence exception and breakpoint resume. By optimizing double-buffer logic and concurrency strategy, improved upload energy efficiency ratio from 3:1 to 10:1, achieved upload success rate of 92.5%, and shortened average closure time of key P0 failures to 0.5 days.

○ Full-scenario ecosystem and plugin system: Built Xiaomi’s next-generation AI super entry based on user data. Wrote 7,000+ lines of full-stack code and developed plugins: ① AI Werewolf: supports users interacting with friends’ digital humans in the same session; ② Life recommendation: integrates Amap POI to realize proactive recommendation and one-click navigation, and expands task flows for Xiaomi “Human-Car-Home” ecosystem, verifying the scalable path from “plugin → action”.

○ Security architecture and privacy compliance: Promoted the implementation of hardware-level TEE encryption link, realized end-to-cloud full-link source-file encryption, and ensured data privacy security at storage and transmission levels in highly sensitive scenarios.

○ Core achievement: Promoted the construction of Xiaomi Auto public opinion and media digitalization system, using LLM + big data architecture to realize precise mining of tens of millions of data and closed-loop real-time PR monitoring and processing.

○ Architecture design and data governance: Led the R&D of “Echo Public Opinion KOL Management Platform”, designed automated data collection and preprocessing system, solved external heterogeneous data synchronization bottlenecks; designed KOL rating model based on multidimensional weighted algorithm, quantified operating indicators (CPM/CPE/CPC) across Douyin, Weibo, Bilibili and other multi-source platforms, and drove media placement strategy from experience-oriented to algorithm-oriented.

○ High-performance system iteration: Deeply participated in the data architecture upgrade of “Echo Public Opinion 2.0”, designed and implemented MySQL + Elasticsearch (ES) heterogeneous data synchronization link; introduced Kafka message queue to realize asynchronous decoupling and peak shaving, supporting hourly incremental updates of tens of millions of indexes; promoted microservice decoupling of system architecture, improving functional module reuse rate by 40% through API orchestration.

○ LLM-driven data mining: For intelligent driving accident scenarios, used SQL complex queries and feature engineering to pre-screen tens of millions of video samples; adopted GPT-4o mini combined with Few-shot Prompting and CoT to optimize labeling logic, collaborated with algorithm team to iterate recall model (Recall), and improved recognition accuracy of key driving accident scenarios to 81.29%.

○ Agentic intelligent tool flow: Self-developed Weibo public opinion automated governance Bot “Heimi Terminator”, built Agent based on DeepSeek-R1, mounted Xiaomi vehicle model knowledge base and PR wording constraints through RAG, realized highly controllable copy generation; connected real-time public opinion API stream, building a second-level automated PR closed loop of “monitoring-analysis-response”.

○ Product design: Led the development of “Yijia” AI development platform, completed requirement research, PRD document writing and prototype design (Figma), performed instruction fine-tuning based on LLaMA2 and StableDiffusionV2 models, supported users without computer background to develop AI workflows in an integrated way, improved enterprise efficiency by more than 30%, and provided highly available privatized AI solutions for 5 leading insurance companies.

○ Full-stack function integration: Led instruction fine-tuning based on LLaMA2 and Stable Diffusion, built 17 high-quality SFT datasets for the financial insurance industry, successfully achieved semantic alignment of models in vertical domains. Defined and implemented more than 30 core components, covering RAG knowledge base, automated data cleaning and image creation modules, and designed and implemented low-code AI workflows.

○ Model privatized deployment: Led the deployment solution design of FSG 13B/33B privatized large models. For the bottleneck of financial long-document parsing, introduced FlashAttention and dynamic VRAM management technology, supporting stable 8K-context inference. Through optimization of the model inference acceleration engine, improved system response speed by 32% and significantly reduced privatized computing power cost.

○ Maintained & optimized the company’s website using React & Django, improving loading speed by 30%.

○ Developed a Content Management API using Node.js & PostgreSQL, enabling automated content updates for company news, artist profile and film details.

○ Implemented SEO & performance monitoring using Google Analytics, enhancing page ranking by 20%.

○ Skill association: Business Research and Software Design.

○ Skill association: Finance and Business.

○ Skill association: Databases and Data Analysis.