PhD · Computer Science · UT Dallas
I build the systems beneath AI search and memory: disk-based vector databases that scale to billions, and retrieval that helps LLM agents remember.
Research
Four first-author systems: two vector-database engines that push billion-scale nearest-neighbor search past the state of the art, and two retrieval systems for LLM agents: multi-hop RAG (UnitRAG) and a multimodal memory system (V-Mem).
A disk-based graph vector-search engine (~6K lines of C++) whose page-node graph structure is co-designed with a storage-aware disk layout, enabling billion-scale search on a single machine with minimal RAM. Adopted by Alibaba's Vector Search team.
Makes graph search explicitly I/O-aware by co-optimizing computation and I/O access: a look-ahead search that adapts across query stages, a priority I/O–CPU pipeline that works cached candidates during I/O waits, and a lightweight in-memory graph index for fast convergence.
A graph-free system for multi-hop question answering that orchestrates atomic LLM tasks over a reusable fact table. It decomposes a question into atomic sub-questions, answers each from retrieved passages, and stores results in a fact table reused across the query, so intermediate reasoning is reused rather than recomputed. No knowledge graph to build or traverse. Highest accuracy among compared methods on three multi-hop benchmarks.
A graph-free, training-free multimodal memory system for LLM agents. It diagnoses two failure modes of similarity-based memory — a cross-modal modality gap and a within-modality similarity–relevance gap — and routes retrieval across a query×target modality matrix. Its index is built with local encoders only: zero LLM tokens, seconds instead of hours.
A taxonomy of agent-memory systems (semantic, entity-centric, episodic, structured) and an empirical analysis of how they are evaluated and where they break.
A PPO agent built from scratch (CNN–LSTM) for active visual tracking, integrating predictive learning with reactive control.
RNNs built from scratch with BPTT and SVD-based orthogonal weight projection; characterized the stability–expressiveness tradeoff.
Prolonging solid-state-drive lifetime — storage-systems work underlying the disk-aware thinking in PageANN and LAANN.
Integrating recent decision models (Jev, Laya) into RAG and agent-memory pipelines for query decomposition, knowledge-graph traversal, and prompt-security guardrails that detect and block insecure prompts.
Publications
Scalable Disk-Based Approximate Nearest Neighbor Search with Page-Aligned Graph
LAANN: I/O-Aware Look-Ahead Search for Disk-Based Approximate Nearest Neighbor Search
V-Mem: Modality-Routed Retrieval for Long-Term Multimodal Agentic Memory
Simpler Tasks, Better Answers: Orchestrating Atomic LLM Tasks over a Reusable Fact Table for Multi-Hop RAG
Augmenting Reinforcement Learning with Visual Attention in Robotics
Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations
Improving the Sustainability of Solid-State Drives by Prolonging Lifetime
Algorithm to Build a Quantum Circuit from a Classical Description of a DFSM
Experience
Vector database systems, LLM retrieval, and deep-learning optimization. First-author work on PageANN, LAANN, and V-Mem.
Knowledge-infused NLP with BERT and LSTM; a real-time Spark Streaming pipeline for tweet ingestion, classification, and geospatial visualization; RL and recurrent-architecture research.
Full-stack iOS application on AWS (API Gateway, Lambda, DynamoDB, S3, Cognito), building REST services end to end.
Shipped three apps to the Apple App Store, with backends on Firebase, Express.js, and MySQL.
Building
Research is one half; the other is putting products in front of real users. A portfolio of apps I have designed and built.
Snap a page and instantly read the Bible verses it references (Recovery Version), in English, 中文, or Español. Built solo end to end — product, engineering, design, and release — from first commit to an approved store release in under two weeks. On-device OCR extracts the references; an optional on-device LLM rescue for OCR-mangled text is re-validated against a deterministic layer, so the model can propose but never decide.
Flutter / Dart · Apple Vision + ML Kit OCR · 132 tests in CI · ~6.6K LOC
Keep the words you hear. NoteWeave turns text from video captions, live speech, and screenshots into searchable notes, organized into folders. Everything runs on-device with no data collection, in English and 中文. Built and shipped solo.
View on the App Store ↗iOS / iPadOS / macOS · on-device OCR + speech
Turn daily goals into a game: track habits and tasks, keep streaks with daily check-ins, and grow a garden of collectible rewards as you make progress, synced across phone and web. Built and shipped solo.
View on the App Store ↗iOS / iPadOS / macOS · cross-device sync
An aviation and aerospace learning platform for K-12 students, parents, and teachers — activities, resources, and tools for educators to find classroom curriculum. Published with Oklahoma State University.
View on the App Store ↗iOS · Education
A "right plant, right place" guide for Oklahoma gardeners: curated recommendations by light, soil, water, and cold tolerance, with fact sheets and shopping lists. Published with Oklahoma State University.
View on the App Store ↗iOS · Education
About
I am a Computer Science PhD student at UT Dallas, working where systems meet AI: how careful system design makes large models faster, cheaper, and more reliable.
Most of my work lives close to the storage layer. My vector-search engines, PageANN and LAANN, treat disk I/O as a first-class design constraint and reach billion-scale nearest-neighbor search on a single machine — one of them adopted by Alibaba's Vector Search team. From there I moved up the stack to retrieval and memory for LLM agents, where V-Mem asks how an agent can remember across text and images without paying for an LLM to build its index.
The through-line is retrieval: finding the right piece of information, at scale, for the model that needs it. Alongside research I ship software — iOS and cross-platform apps — because building for real users keeps the research honest.