Wei-Ting Wu

Wei-Ting Wu

Taipei, Taiwan

I focus on the alignment problems that emerge when multi-agent systems run for hours, coordinate across thousands of steps, and recover from partial failure.

Current focus

Alignment over time.

A multi-agent system can begin with a clear objective and still drift as plans change, agents exchange partial context, and small errors compound over a long run.

The problem is not only whether each agent is capable. It is whether the system can preserve intent, coordinate decisions, and recognize when its own trajectory has become unreliable: goal alignment, coordination, and recovery.

I study and build the mechanisms that keep long-running agentic systems inspectable, steerable, and aligned with the task they were given.

Projects

AI infrastructure

share-house

Run private AI services on shared infrastructure without exposing one team’s data or workloads to another — and without rebuilding the same environment for every team.

  • Multi-tenant inference scheduling
  • Tenant isolation
  • Edge-node orchestration

long-horizon agents

digs

Hand long, multi-step security investigations to agents without supervising every move.

  • Offensive and defensive agent roles
  • Shared task state and role delegation
  • Long-horizon planning and tool use
  • Retry and failure recovery

observability

horus

Find where an agent run went wrong without reading scattered logs.

  • OpenTelemetry traces
  • Langfuse and Arize Phoenix
  • Model, tool, and handoff tracing
  • Latency, cost, and failure attribution

on-device AI

WebGPU LLM

Use an LLM without uploading private data or waiting on a remote server.

  • SFT and GRPO post-training
  • PEFT under limited compute
  • Model quantization
  • On-device inference through WebGPU
  • Presented at JSDC 2025

research agent

empaithy

Run more user interviews without adding more interviewers, and turn every conversation into usable research output.

  • Real-time multi-turn voice interviews
  • OpenAI Realtime
  • Transcription and summarization
  • Structured insight extraction

multi-agent simulation

WAVE

Before executing a decision, see how different stakeholders may understand it, respond, and influence one another.

  • Personas distilled from real interviews
  • Multi-agent influence across multiple rounds
  • Complex-system modeling for dynamic game-equilibrium simulation