WEI Lingfeng

Leading frontend development and participating in multi-agent architecture design, building an AI-driven drug discovery platform for scientific collaboration. Designed and implemented Copilot-style multi-user document collaboration mechanism and agent-based assistant framework to support research workflows.

Independently designed and implemented the inference sampling module of a reinforcement learning framework, supporting advanced features such as multi-turn conversations, asynchronous batch processing, customizable agents, and vision-language model integration. This framework supported the training of INFLY-RL-SQL-32B, which achieved 2nd place on Bird-Bench in April 2025 and is currently being used to train vision-language models for Android World.

Led the design and implementation of function calling and structured output protocols. Proactively collaborated with business teams to unify product interfaces across use cases, forming a unified API layer used by internal and external stakeholders. Helped secure international clients through platform openness and reliability; laid technical foundations for SFT pipelines, reinforcement learning training, and agent orchestration within the company.

Designed and implemented INF's online model service open platform, focusing on key features such as API key management, gateway, recharge, billing, invoice details, and interface standardization.

Developed the company's early-stage chatbot application Inf-Chat. Played a key role in INF's first version of foundational model(66B) training and release, large model registration, financing, and business promotion.

Collaborated with drug R&D teams, used Figma to design prototypes and developed prototypes for multiple drug R&D projects: drug discovery platform, drug design platform, compound management platform, etc.

Project Introduction: A cloud drive for designers, providing excellent preview functionality for numerous file extensions and quick collection tools with Chrome extension. With MuseDAM, it's easy to classify, upload, download, and share files. Achievement: Efficient, high-quality styled architecture with excellent development experience. Unified style specification for code and design drafts (Figma), vite development tools, unit testing with jest and end-to-end testing with playwright, etc. Achieved 5000 registered users and 1500 weekly active users one month after release.

Main functionality: A dashboard that updates frequently in real-time, displaying current status, historical trends, key parameters, and alarms of nuclear power units. Achievement: Thanks to the Vuetify-based admin template, the project was quickly migrated from Django SSR to Django REST framework + Vue architecture.

Introduction: Fitting the sparse measurement physical distribution (56*50 array) of the reactor's 3D core to a denser distribution (56*157*289 array). Achievement: Achieved excellent algorithm performance in terms of speed and accuracy.