Information Retrieval · ML Systems · Backend Reliability

Computer Science undergraduate at Tunghai University (B.S. expected 2028) in Taichung, Taiwan. I build measurable systems that connect research ideas with deployable software—from embedding compression and ranking diagnostics to failure-aware backend infrastructure.

My work emphasizes reproducible evaluation, realistic deployment constraints, observability, and transparent reporting, including results that do not support the original hypothesis.

• Current research: frozen-index correction for compressed retrieval and a focused short-paper manuscript in preparation
• Seeking: internships in information retrieval, ML systems, search/RAG, and backend engineering
• Contact: [email protected]

A reproducible benchmark, frozen-index retrofit study, and deployable retrieval system for compressed embeddings and Faiss IVF-PQ search.

• Benchmarked PQ, OPQ, and GPU IVF-PQ across FiQA, SciFact, and a deterministic 1M-passage MS MARCO corpus, covering 40 compression and search configurations.
• At M=96, nprobe=64, IVF-PQ retained 92.1% of exact Recall@10 at 13.01× serialized deployment compression; Faiss-only search on a Tesla T4 reached 0.072 ms P95 latency and 14,619 QPS.
• Developed Retrieval-Aware Residual Subspace (RARS), a rank-16 int8 sidecar that corrects candidate scores without rebuilding the frozen IVF-PQ index or rewriting its PQ codes.
• On the full 1,000-query held-out MS MARCO split, RARS improved Recall@10 from 0.6833 to 0.7073 (+0.0240; 95% CI [+0.0105, +0.0378]). After excluding 137 queries used in earlier exploratory work, the remaining 863 prior-unseen queries retained a positive Recall@10 gain of +0.0168 (95% CI [+0.0029, +0.0303]).
• Storage-matched comparisons on a corpus-restricted TREC DL analysis and the 3,452-query official BEIR NQ test did not establish RARS superiority over PCA. These results are retained without test-time retuning and motivate a separately versioned boundary-loss feasibility study.
• Packaged the work with versioned artifacts, SHA-256 verification, paired-bootstrap evaluation, FastAPI search and batch endpoints, Docker, tests, CI, and paper-ready CSV/LaTeX tables.

Latest research branch · Short-paper draft

A production-minded backend focused on concurrency, durable event delivery, security, observability, and failure recovery.

• Implemented JWT refresh-token rotation and revocation, multi-device sessions, payment idempotency, optimistic locking, and resource-ownership checks.
• Built a transactional outbox with PostgreSQL SKIP LOCKED claims, processing leases, Kafka retry/DLT handling, idempotent consumers, and administrative replay.
• Added Prometheus, Grafana, Alertmanager, security audit signals, and 40 Testcontainers tests across PostgreSQL, Redis, and Kafka.
• Validated the catalog read path locally with 750,000 requests, 2,499.91 requests/s, 0.84 ms P95 latency, and no observed application-side 5xx responses. This was a Docker-on-macOS soak test, not a production SLA or cloud-capacity claim.

A Linux runtime-extension project in C.

• Loads payment and discount plugins from configuration using dlopen, dlsym, and shared-library interfaces without recompiling the host.
• Validates required symbols, records transactions, and includes automated extension and broken-plugin failure demonstrations.

An exploratory machine-learning pipeline for dialysis time-series records.

• Processed 6,279 sessions into 328 statistical and temporal features, then trained a top-30-feature XGBoost model with group-aware cross-validation.
• Reported approximately 0.86 mean balanced accuracy and 0.87 mean F1-score on the evaluated dataset.
• The project demonstrates leakage-aware evaluation and feature engineering; it is not clinical validation or a diagnostic system.

• Retrieval and ML: Python, PyTorch, Sentence Transformers, BEIR, Faiss, IVF-PQ ADC, PQ/OPQ, scikit-learn, XGBoost
• Backend and data: Java, Spring Boot, FastAPI, PostgreSQL, Redis, Kafka, JPA/Hibernate
• Reliability and delivery: transactional outbox, idempotency, optimistic locking, Testcontainers, Docker, GitHub Actions
• Observability: Prometheus, Grafana, Alertmanager, Micrometer
• Systems: C, C++, Linux, Bash, Git, Make, shared libraries, dynamic linking

• Measure retrieval quality, storage, latency, and throughput together.
• Separate exploratory findings from confirmatory evidence and preserve negative results.
• Design backend systems for retries, duplicate delivery, concurrency, recovery, and observability.
• Keep experiments, artifacts, and performance tests reproducible.

• Email: [email protected]
• Location: Taichung, Taiwan
• GitHub: ravan-chuang