An exact vector-search engine built from the CUDA kernel upward through C++20, Java JNI and a hardened Spring Boot API. The system keeps indexes resident on the GPU, supports FP32 and FP16-storage modes and includes a reproducible benchmark harness.
Measured evidence: 16.7× lower P50 latency than the project’s scalar CPU reference on 500K × 128 vectors · 1,641–1,662 FP32 QPS at batch 32 · 48/48 benchmark cases passed
Architecture · Benchmark methodology · v1.0.0 release
A Java workshop that develops a distributed transactional key-value store in six deliberate stages: hybrid logical clocks, MVCC, distributed routing, transaction records and intents, clock-uncertainty restarts and serializable conflict prevention.
Engineering focus: transaction ownership · snapshot isolation · write skew · uncertainty windows · read restart · intent resolution
| Project | Engineering problem | Status |
|---|---|---|
| AutoMQ #3493 | Prevent the metrics reporter from starting on controller-only nodes | Merged |
| Trino #30973 | Coordinate query failure with transaction-commit ownership | Review |
| Fluss #4263 | Make cached RPC connection identity endpoint-aware | Review |
| AutoMQ #3579 | Add optional authentication to the built-in Prometheus endpoint | Review |
| Fluss #4230 | Support custom Paimon lake-table paths in Spark reads | Review |
Backend — Java 8–21 · Spring Boot · REST · gRPC · concurrency · asynchronous processing
Data and messaging — PostgreSQL · MySQL · MongoDB · Redis · Kafka · event-driven systems
Reliability and performance — idempotency · rate limiting · circuit breakers · caching · SQL tuning · JFR
Platform and operations — Docker · Kubernetes · AWS EKS · Jenkins · CI/CD · CloudWatch · Splunk · Dynatrace
Validation — JUnit · Mockito · deterministic race tests · benchmark design · production telemetry