Yuchen Liu

Search / AI infrastructure engineer

I build production-scale search and AI infrastructure for model research: streaming indexing, hybrid retrieval, and reproducible LLM/RL tooling.

My work sits between research and production systems: company-scale indexing and retrieval infrastructure, nanoRL for hackable RL experiments, and co-first-author research on LLM architecture and mechanistic interpretability.

GitHub X LinkedIn Resume
01 · nanoRL Open-source RL training framework for objectives, rollouts, vLLM-backed serving, and reproducible experiments.
02 · Search infra Streaming indexing and Lucene-native hybrid retrieval systems with Spark, Kafka, Iceberg, and dense embeddings.
03 · Research systems LLM mechanisms, VLM formalization, and agentic medical imaging work grounded in experiments and runnable tooling.

Selected work

Selected AI systems projects

01

nanoRL

Led an open-source RL training framework inside nanochat for objective design, rollout experiments, reproducible ablations, and vLLM-backed inference serving.

  • RL training
  • vLLM
  • Rollouts
  • Reproducibility
 02

Search and data infrastructure

Built production indexing and hybrid retrieval infrastructure with Kafka, Spark, Iceberg, Lucene, BM25 retrieval, and dense embeddings.

  • Spark
  • Kafka
  • Iceberg
  • Search
 03

Research-grade model systems

Co-authored LLM architecture, VLM formalization, and agentic medical imaging work, with an emphasis on mechanisms, ablations, and tooling.

  • LLMs
  • VLMs
  • Planning
  • Medical AI

Experience

Production systems with research taste

Software Engineer, Search Infrastructure

DoorDash · Aug 2025 - Present

Owned core components of a next-generation search ingestion platform, re-architecting indexing write paths across production search stacks, sharded indexers, large document corpora, and tens of TB of index storage.

Designed streaming incremental indexing with Kafka, Iceberg, Spark, watermarks, and exact cutover offsets; productionized Lucene-native hybrid retrieval that fuses BM25 with dense embeddings.

Software Engineering Intern

Penn Medicine TissueLab · Oct 2024 - May 2025

Developed a FastAPI/PyTorch microservice platform and event-driven DAG workflow engine for medical image analysis, supporting natural language-driven orchestration over PyTorch-based segmentation and classification tools.

Software Engineering Intern

Information Technology of CAS · Apr 2024 - Aug 2024

Designed Kafka and Redis-backed event streaming services for IoT monitoring, reducing p95 API latency from 2s to 200ms while scaling distributed Spring Boot services.

Research

Papers that inform the systems work

Bank of Values architecture figure

EMNLP 2026 · Under review

Do Value Vectors in Deep Layers Need Context from the Residual Stream?

Co-develops the hypothesis that late-layer value vectors can encode context-free token information, with controlled architecture ablations across 21 benchmarks.

Vision-PDDL planning task pipeline

ECCV 2026 · Under review

Vision Language Models Cannot Plan, but Can They Formalize?

Studies whether VLMs can turn visual observations and natural language goals into solver-executable PDDL problem files, with formalization and solver-based correctness checks.

Agentic medical imaging workflow diagram

Nature · Under review

A co-evolving agentic AI system for medical imaging analysis

Co-develops agentic medical-imaging workflows for tool selection, planning, knowledge updates, and human-in-the-loop co-evolution.

AI systems projects

Small stacks for serious experiments

nanoRL

RL training framework for objective design, rollout experiments, reproducible ablations, and vLLM-backed serving.

 nanochat

Hackable pretraining and chat stack for architecture ablations and reproducible model experiments.

 llm-interp

Lightweight interpretability scripts for model circuit research and reproducible LLM behavior studies.