max xiong
current studying cs & math at duke. talk to me about everything AI, research, startups, building!
oh, i also trained an llm to talk like me. try it out (soon) :)
experience
Founding Engineer #4
-
Founders Fund, 10M+ ARR
Research Intern
Microsoft AI
RL + SFT for reasoning LLMs over detection of AI-generated text. Currently achieving SOTA performances, publishing paper this January.
Organizer
HackDuke
Organizing Duke University's largest hackathon, managing logistics and operations for 500+ participants annually.
Research & Teaching
Duke University
Research in ML for Ecology, AI Agents, and Reasoning. TA for Data Structures & Algorithms.
Researcher
Rutgers University
Contributing to research initiatives in recommender systems and ML.
projects
HackDuke Portal
Duke University's Largest Hackathon
Full-stack web application managing applications and registration for HackDuke, serving 500+ participants annually. Features Auth0 authentication, dynamic form system with file uploads, AWS S3 integration, and production-ready Dockerized infrastructure with comprehensive testing and monitoring.
Popstar: Agentic RL Training
Weights & Biases Hackathon
Winner: Best Use of Weights & Biases. Developed a self-improving AI agent that learns to play Kirby's Dream Land using reinforcement learning and LLMs. Used evolutionary algorithms and LLMs to propose novel intrinsic rewards and training algorithm modifications. Built with OpenEvolve framework, demonstrating autonomous learning through PPO with raw pixel inputs.
Jamfusion
HackHarvard 2025: Compile the Decade
AI-powered collaborative music creation platform enabling users to explore and remix global musical traditions through an interactive flow diagram. Features LLM-powered recommendations, speech-to-graph creation, and cross-cultural fusion capabilities across 71+ global cultures.
Jasma
HackGT 12: Midnight at the Museum
AI-powered developer tool that captures, validates, and shares bug fixes across a collaborative knowledge base. Created a complete end-to-end system with repo snapshots, semantic retrieval using pgvector, and automated fix validation to eliminate repetitive debugging.
Watson: AI Agents for Crime Scene Investigators
Y Combinator AI Agents Hackathon
Built an MCP server with Dedalus Labs AI to let an LLM call an external custom Python compiler agent tool for automatic graph generation with Matplotlib. Developed a FastAPI backend and Next.js frontend with Supabase integration for storing and retrieving unstructured case files.
Seek Agent
tryseekagent.com
Developed a real-time stock screener web app streaming market data for 500+ tickers. Integrated Financial Modeling Prep API for financial statements/ratios and used LangChain to build agentic tools surfacing value-investing insights. Released beta with 100+ active users, demonstrating scalability in API integration and real-time data delivery.