Tao Shen

AI / LLM Engineer · Post-Training (SFT, LoRA, Model Merging) · Distributed & Federated Training AI Agents

Hangzhou, CN.

About

Highly accomplished AI/LLM Engineer and Ph.D. in Computer Science with 20+ publications and 1,470+ citations, specializing in distributed optimization, federated learning, and model merging. As the founder of Democra AI, independently shipped three innovative AI products, demonstrating end-to-end expertise from research to product delivery. Proficient in PyTorch, LLM post-training (SFT/LoRA/PEFT), and model merging, driving AI-native workflows and architectural ownership.

Work

Democra AI, Inc.

|

Founder & CEO

Delaware, USA, US

→

Summary

Led Democra AI, a one-person startup, from vision to product, shipping three innovative AI solutions and establishing a robust tech stack.

Highlights

Solely founded, designed, and engineered Democra AI, successfully shipping three distinct AI products (0→1) within one year and establishing a reusable tech stack and release workflow.

Developed and launched CiteTrack, a native iOS/iPadOS/macOS/visionOS app for researchers to track Google Scholar impact via an AI agent, successfully passing Apple review as a single universal app.

Created and deployed AIR (air.democra.ai), a bilingual web platform that assesses personal AI-replacement risk by synthesizing public data from MIT/McKinsey/WEF/PwC into industry-level views.

Engineered Candy Shop (candy.democra.ai), a two-sided marketplace enabling domain experts to package and offer their expertise as invokable AI agent skills.

Pioneered an AI-native toolchain, including HuggingClaw (53 stars), claude-workflow-viz, skill-to-workflow, and Achilles, enhancing development efficiency.

Alibaba DAMO Academy

|

Research Intern | Distributed Machine Learning

Hangzhou, China, China

→

Summary

Contributed to the development of distributed ML and federated learning systems at Alibaba DAMO Academy, optimizing performance for large-scale, heterogeneous data environments.

Highlights

Contributed to the design of distributed ML / federated learning systems, supporting large-scale, highly heterogeneous client data in production settings.

Optimized aggregation and communication strategies, significantly reducing communication overhead while preserving model quality (metrics not disclosed for compliance).

Helped build Luoxi, an open-source device-cloud collaborative model platform (92 stars), enabling adaptive aggregation across heterogeneous devices.

Zhejiang Lab

|

Research Intern | Federated Learning

Hangzhou, China, China

→

Summary

Conducted research on federated learning and distributed AI systems, enhancing training efficiency and developing a widely-cited mutual learning paradigm.

Highlights

Researched federated learning and distributed AI systems, running controlled experiments across heterogeneous settings to improve distributed-training efficiency.

First-authored 'Federated Mutual Learning' (arXiv:2006.16765), a widely-cited paradigm with 302 citations, enabling client models to retain individuality while learning shared knowledge through mutual distillation.

Zhejiang University (College of Control Science & Engineering)

|

Research Assistant | Numerical Optimization

Hangzhou, China, China

→

Summary

Developed and validated a graph-based distributed method for power-flow computation, avoiding reliance on complex matrices and demonstrating scalability.

Highlights

Developed a graph-based distributed method for power-flow computation, eliminating reliance on BIBC/BCBV matrices and modeling SDP/SOCP convex programs with YALMIP.

Validated convergence and scalability of the developed method on large systems.

First-authored 'A Graph-based Power Flow Method for Balanced Distribution Systems' (Energies, 2018), accumulating 115 citations.

Education

Zhejiang University

Hangzhou, China, China
→
Ph.D. Computer Science

Courses

Distributed Optimization

Federated Learning

LLM Post-Training (LoRA / Model Merging)

Towards Adaptive Federated Learning Methods in Distributed Heterogeneous Environments (Dissertation)

Zhejiang University

Hangzhou, China, China
→
M.S. Control Science and Engineering

Courses

Numerical Methods

Optimization

Power Systems

China University of Petroleum

Qingdao, China, China
→
B.S. (Top-tier Class) Automation

Courses

Neural Networks

Control Systems

Soft Measurement of Reaction Concentration in a CSTR system (Capstone)

Publications

Will LLMs Scaling Hit the Wall? Breaking Barriers via Distributed Resources on Massive Edge Devices.

Explores challenges and solutions for scaling Large Language Models on distributed resources and massive edge devices.

Fedeve: On Bridging the Client Drift and Period Drift for Cross-device Federated Learning.

Research on bridging client drift and period drift in cross-device federated learning.

Knowledge-empowered, Collaborative, and Co-evolving AI Models: The Post-LLM Roadmap.

Engineering

Outlines a roadmap for future AI models that are knowledge-empowered, collaborative, and co-evolving beyond current LLMs.

An Adaptive Aggregation Method for Federated Learning via Meta Controller.

ACM MM

Proposes an adaptive aggregation method for federated learning using a meta-controller.

Federated Graph Learning-A Position Paper.

A position paper discussing federated graph learning.

Federated Mutual Learning.

Introduces a paradigm where client models retain individuality while learning shared knowledge through mutual distillation.

A Graph-based Power Flow Method for Balanced Distribution Systems.

Energies

Presents a graph-based distributed method for power-flow computation in balanced distribution systems.

Languages

English

Chinese

Skills

Programming & Frameworks

Python, PyTorch, FastAPI, Docker, C/C++, MATLAB.

LLM & AI Engineering

LLM post-training (SFT / LoRA / PEFT), Model Merging (model fusion / MOE-LORA), Federated Learning (FedLLM, federated distillation, graph FL), Hugging Face (Transformers / PEFT), Flower (FL framework), LLM fine-tuning, Continual Learning, Catastrophic-forgetting mitigation, AI-agent & skill design, Claude Code / AI-coding workflows, MCP.

Optimization & Data

Distributed Optimization, Numerical & Convex Optimization (SDP / SOCP), Heterogeneous-data adaptive aggregation, Data Analysis.

Web & Mobile Development

Web Full-stack (Next.js / React / Tailwind), iOS / macOS Native (SwiftUI / WidgetKit), Cloudflare / Vercel, SQL.

Working Style & Methodologies

0→1 Product Delivery, Bridging Academic Research and Industrial Deployment, MVP Development, Launch & Iterate Cycle.

Projects

SenseVoice-LoRA-MPS

→

Summary

Developed an end-to-end LoRA (PEFT) fine-tuning pipeline for the SenseVoice speech model on Apple Silicon (MPS).

FlowerTune

→

Summary

Developed a cross-domain federated fine-tuning of LLMs with Cambridge Flower Labs.

Parameter-Efficient Fine-Tuning & Model Merging

→

Summary

Contributed in-depth to method design and experiments for a series of model-merging / PEFT works.

HuggingClaw (free agent deployment on Hugging Face Spaces)

→

Summary

A free agent deployment tool on Hugging Face Spaces.

Mobile Intelligent Dataset (Alibaba Tianchi)

→

Summary

A mobile intelligent dataset from Alibaba Tianchi.

FEMNIST for PyTorch (Dataset)

→

Summary

A PyTorch implementation of the FEMNIST dataset.

simple-FL (research training framework)

→

Summary

A research training framework for federated learning.

Federated-Learning-FAQ

→

Summary

An introductory tutorial on federated learning, companion to a Momodel seminar with 22,000+ replay views.