AI Age Research and Development Engineer (internship)
Internet Recreation (Shanghai) June 2026 August 2026 Intern Rewind
- Project background and objectives:Develop an Agent-based UI Project Auto-Generation System, reduce duplication of construction work when converting UI designs to online projects, reduce process dependence on personal development experience, and explore the role of Agent in a more complete UI workflow.
- UI Agent Workstream:Design and implement a hybrid structure independently, combining the definitive Age Workflow with the OpenAI Actors SDK based autonomous nodes. System can be generated from the Figma/PSD script NeoX
.uiprefabProject, while supporting generation from zero and blueprint-based. The process package has about 50 nodes, 26 of which are reused between different processes. LLM only handles tasks that do require model judgement, such as structural extrapolations, selection of materials, and determines that sex workflows are responsible for the bottom-out; multi-modular retrieval and project specifications allow the results to follow established engineering specifications and re-use existing controls as much as possible. Blueprint models can re-use historical engineering. I worked with the GUI designer to organize about 10 business blueprints covering common interfaces like weekly sign-ins, shop windows, etc. Part B/C UI missions have been reduced from about 0.5 to 1.5 person days to 30 minutes, and 4 generation projects have been online. - UI Agent evaluates and returns:Create a three-tier validation process for "inspiring rules screening, Ground Truth comparison, manual final review" to make Agent's results more stable and more easily validated. Use AI to dig up structural clustering and sub-prefab candidate models from existing projects, then develop standard calibrations according to UI and add these models to the workflow through multiple rounds of Skill Demo. For more than 10 real UI samples, continuous iteratives are checked with an inspirational rating and editor, supplemented by regression tests for blueprint generation and base capacity from zero. In a blueprint intermediate migration, field and script omissions were detected; these omissions could lead to a distorted overall picture ratio.
- Internal knowledge base:Try different knowledge modelling and retrieval options for knowledge in complex types of internal documentation and game fields. For documents that are classified, in multiple file formats, a unified structured factual layer is created using Docling, which complements chart information, then integrates vector search, BM25 and PageIndex style structured search. For game knowledge, about 20,000 physical nodes and 30,000 program relationships are extracted from the source code and configuration table and organized into JSON Wiki to address the problem of planning for expired documents and conflict of content. On this basis, the LLM Wiki style and query rewriting is used to construct the Agenic RAG to achieve the recall of all questions on the test set.