Build the Base.Power the Intelligence.

Own AI Agent product planning and roadmap, define product positioning and target customers, continuously track technical developments in the Agent space, and deliver competitive analysis and product iteration direction.

Own planning and roadmap for Token/MaaS products, define product positioning, target customers, and business models, continuously track technical developments in the field, and deliver competitive analysis and product iteration direction.

Develop the traffic plane and control plane of the LLM inference gateway: multi-provider and multi-model access, routing and scheduling, rate limiting and circuit breaking, retries and failover, key and quota management, and SSE/streaming forwarding — ensuring latency and availability under high concurrency and long-lived connections.

Research and implement core algorithms for LLM inference acceleration — model distillation, PTQ/QAT, structured/unstructured pruning, and sparse inference — systematically reducing inference costs.

Design and build cloud-based inference service platforms for large-scale language and multimodal models, supporting mainstream open-source models (LLaMA, ChatGLM, Mistral series) for online/near-online inference deployment.

Test architecture design and development for LLM inference services — covering functional correctness, performance stability, and user experience reliability, ensuring high-quality delivery of inference APIs and platform.

Design and develop frontend pages for web products, delivering a high-quality user interaction experience.

Server-side architecture design and R&D of core business systems — leading technical reviews and core module coding, ensuring high performance, high availability, and horizontal scalability.

Own test development for the company's AI products and platforms, participate in requirement reviews, and create test strategies and test plans.

Responsible for production service and AI infrastructure stability — container orchestration (Kubernetes), distributed storage (Ceph/MinIO/Redis), high-performance networking (RoCE/InfiniBand) — ensuring 7x24 high availability of core services.

Lead the build-out and continuous improvement of MLPS 2.0, commercial cryptography assessment, ISO 27001, and the information security management system; lead annual assessments and surveillance audits, and liaise with third-party assessors;

Design and implement core data models and service interfaces for tasks, sessions, messages, and execution context.

Design memory data models, storage structures, and access interfaces for different scopes such as personal, project, and workspace.

Build dataset import, cleaning, versioning, quality analysis, and lineage management so that training and evaluation data is traceable.

Own the overall goals of the data engineering team: plan the AI data asset system (full pipeline of collection, cleaning, processing, and evaluation) and deliver high-quality datasets to both the pre-training and post-training consumption lines.

Build and maintain LLM corpus pipelines: crawling, parsing, deduplication, desensitization, and format standardization of text sources such as web pages, books, papers, and code.

Build multimodal data pipelines: collection, decoding, splitting, deduplication, and metadata extraction from image, video, and audio sources.

Develop the data service platform: dataset storage, versioning, lineage tracking, and retrieval services, providing unified data consumption interfaces (API/SDK).

Own the overall goals of the post-training team: plan technical routes such as SFT (Supervised Fine-Tuning), RLHF (Reinforcement Learning from Human Feedback), and DPO (Direct Preference Optimization) to turn base models into deliverable product models.

Research and implement LLM post-training algorithms: SFT data recipes, RLHF/DPO/preference optimization, process reward models (PRM), and other frontier methods.

Research post-training algorithms for multimodal generation models, covering image generation, video generation, and joint audio-video generation; continuously improve generation quality, instruction following, aesthetic quality, temporal consistency, and controllability.

Build the model evaluation system: combining automated benchmarks, human evaluation processes, and online metrics, covering both LLM and multimodal models.

Build and maintain post-training pipelines: SFT/RLHF training frameworks, distributed training jobs, checkpoint-resumed training, and experiment tracking and configuration management.