Systems
Selected systems spanning context infrastructure for agents, production graph systems, tensor-centric execution, and earlier database/graph research.
These systems trace a line from distributed query and graph execution, through production graph services and accelerator-aware execution, to self-improving context infrastructure for agents. Each entry summarizes the problem and core system thesis; the detail pages connect that thesis to architecture, research threads, evidence, and system boundaries.
Current
Agents are limited not only by model capability, but by the external information environment available to them and the task-specific context constructed from it. AutoIA treats both as optimizable state: an inner loop improves environment-specific retrieval pipelines from task-level evaluation, while an outer loop uses persistent failures to change data integration, organization, indexing, and storage. It is the systems platform for my current self-improving context-infrastructure agenda.
Huawei Systems
GES is a production graph database service for high-concurrency interactive workloads, where throughput, latency, extensibility, and operational robustness must hold together. The research line combines composable service architecture, factorized execution, dynamic graph storage, and incremental processing. It reached #1 on both LDBC SNB Interactive declarative and imperative tracks, with the declarative result reporting more than 3,000× the throughput of the previous #2 system.
TQEX asks how SQL and graph systems can benefit from rapidly evolving accelerators without becoming hardware-specific codebases. The line uses tensor runtimes as a portable execution layer, then closes the mismatch between irregular data workloads and uniform tensor operations through compact representations, workload-specific operators, batching, compression, and out-of-memory execution. It spans SQL, graph queries, graph processing, and graph search.
Earlier Research
These systems established the database and graph-processing foundations behind my later work: SeccoSQL separates communication from local computation in distributed query plans; DISC recasts graph analytics as relational execution; Crystal computes compressed subgraph results directly. Together they explore how changing the execution abstraction exposes better optimization opportunities and controls intermediate-state growth.