Steve Kuo
Steve Kuo

Steve Kuo

I'm an undergraduate at UC Berkeley studying Data Science and Mathematics. I'm a software engineer and AI researcher, currently splitting my time between Virtue AI as a Research Scientist Intern and Workday as a Software Engineer Intern. Previously I was in the Founders Inc Canopy cohort at Fort Mason, San Francisco, building HVAC Smart Glasses. My interests span physical AI, robotics, reinforcement learning, AI engineering, low-level systems, and the safety and evaluation of frontier agents.

steve.kuo@berkeley.edu  ·  github  ·  linkedin  ·  twitter

News

  • Jun 2026 Started as Research Scientist Intern at Virtue AI, building automatic red-team agents for frontier LLMs.
  • May 2026 Started as Software Engineer Intern at Workday, working on test infrastructure automation for the payroll team.
  • Apr 2026 AWS Builder Loft: 1st place overall + three sponsor tracks.
  • Mar 2026 AWS GenAI Hackathon: 1st place, Aerospike track. NVIDIA Spark: 1st, Nemotron track.
  • Feb 2026 Authored AgentRearrange: A Technical Writeup for Swarms.ai, on multi-agent task decomposition. [pdf]

Experience

  • Research Scientist Intern, Virtue AI .

    Working on agentic red-teaming for frontier LLMs on top of DTAP (DecodingTrust-Agent Platform). Generated 500+ adversarial datasets across 10 risk categories, achieving >50% ASR (attack success rate) against Opus 4.7, Llama, and DeepSeek.

  • Software Engineer Intern, Workday .

    Test infrastructure for the payroll team's CI pipeline. Built an internal Python service wrapping the payroll integration test runner with automatic retry, using stacktrace-signature matching to skip retries on genuine bugs. Cut manual re-runs on the nightly payroll suite by ~60%, saving ~10 engineer-hours/week on triage. Consolidated 4 service-specific retry configs into a single YAML schema validated in CI.

  • Software Engineer Intern, Fetch.ai .

    Shipped Python/FastAPI microservices on Fetch.ai's platform powering thousands of developers. Built Docker CI/CD for signup, OAuth, and microservice provisioning, cutting deploy time by 60%. Engineered Elasticsearch + semantic embedding RAG with sub-100ms p99 retrieval at platform scale. Designed multi-agent orchestration via context engineering, cutting LLM call latency by 35%.

  • Open Source Contributor, Swarms.ai .

    Open source contributor on the enterprise multi-agent framework (5.2k+ GitHub stars; Python, FastMCP, Rust, TypeScript). Shipped 200+ merged PRs across orchestration, API, and platform. Implemented research-paper architectures (Mixture of Agents, Reflexion, Self-Consistency) into core. Built core orchestration components powering production agents in finance, healthcare, and legal. Wrote the infrastructure for an agent marketplace hosting 10K+ deployed agents handling $100K+ in transactions.

  • Software Engineer Intern, Northrop Grumman .

    Wrote Bash scripts to verify software integrations for airborne communications, cutting test time by 50%. Built C++ modules for real-time signal handling and telemetry used in flight systems. Automated verification with Ansible to interface with firmware, improving repeatability and traceability.

Research

  1. [1]

    Agentic Red-Teaming on DTAP (DecodingTrust-Agent Platform). Virtue AI , 2026.

    Extending the DTAP benchmark with new sandboxes (Gmail, Salesforce, Slack-MCP, GitHub-MCP) and building adversarial prompt pipelines that stress-test frontier coding agents (Claude Code, OpenClaw).

  2. [2]

    AgentRearrange: A Technical Writeup. Swarms.ai , 2024.

    Multi-agent task decomposition and reordering across agent graphs. Patterns from working on the orchestration framework.

    [pdf]

Projects

  1. [1]

    Overflow , 2026.

    Multi-agent RL environment for autonomous vehicle fleet oversight, built on Meta's OpenEnv. 1st place overall at the Founders Inc hackathon.

    [writeup]

  2. [2]

    IntentOverflow , 2026.

    A two-sided marketplace for AI-era bugs. Vibe-coders post broken AI-generated projects with bounties; skilled devs fix them live in an in-browser Cursor-style IDE with AI assist. Deployed with ~100 users.

    [writeup]

  3. [3]

    HVAC Smart Glasses , 2026.

    AR overlay for HVAC technicians. Real-time equipment recognition, schematic recall, and guided repair flows. Offline-first; hands stay on the tools.

    [writeup]

Writing

  1. [1]

    Patterns for Reliable LLM Agents. Personal blog , Dec 2025.

    Practical patterns for agents that don't fall apart in production. Structured output, retries, retrieval, traces.

    [post]

  2. [2]

    Why I Think About Markets. Personal blog , Oct 2025.

    Markets as the most honest real-time feedback system for testing ideas under uncertainty.

    [post]

Hackathons

21 wins across the AWS, NVIDIA, ByteDance, YC, and university circuits.

  • AWS Builder Loft. 1st place overall + Interconnected, TrinityOS, Mindra tracks.
  • AWS GenAI Hackathon. 1st place, Aerospike track.
  • NVIDIA Spark Hackathon. 1st place, Nemotron track.
  • Pokee AI · Dataclaw Hackathon. 1st place overall.
  • NightHack 2. 1st place overall.
  • SJSU Startup Pitch Jam. 1st place overall.
  • SJSU Tractionathon. 1st place overall.
  • Silicon Valley Innovation Challenge. Winner.
  • AWS Builder Loft (Round 2). 2nd place, TestSprite, Minimax, Neo4j tracks.
  • Swarms.ai Hackathon. 2nd place overall.
  • CalHacks. 2nd place, Best Use of AgentVerse.
  • YC Agent Jam. 2nd place, Best Use of Nivara.
  • DigitalOcean × MLH Hackathon. 2nd place overall.
  • ByteDance Hackathon. 2nd place.
  • AWS Builder Loft. 3rd place, Apify track.
  • Frontier Tower · Korean Investment Challenge. 3rd place.
  • SCEHacks. Cybersecurity track winner.
  • StanfordXR. Best Design & Storytelling.
  • Minimax @ Entrepreneur First. Multimodal Award.
  • SFSU Hackathon. Best Use of Vulture.
  • Total Recall Agents Hackathon. Best Use of Dify.

Miscellaneous

Classical violin (15 years), lacrosse, and wrestling. I think the discipline that compounds quietly is what makes the visible stuff possible.