Projects

Technical Implementation

• Built a FastAPI agent with ReAct planning, intent routing, and a confidence-driven decision policy orchestrating four specialized skills for local medical QA, web search, medical-record insertion, and field-level record retrieval.
• Built a local medical RAG pipeline over 72,877 retrieval chunks from 11,274 XML files and 16,407 question-answer documents using contextual query rewriting, Milvus vector search, BM25 hybrid retrieval, and reranking.
• Designed Redis-backed short-term memory with sliding-window retention and rolling LLM summarization.
• Built long-term memory with LLM-based extraction, skip / merge / create update decisions, Redis record storage, and Milvus semantic recall for personalized cross-session responses.
• Developed a PostgreSQL medical-record module with two normalized tables and three custom indexes for structured insertion and field-level retrieval of allergies, medications, symptoms, diagnoses, procedures, and vitals.
• Implemented asynchronous, memory-aware agent endpoints for concurrent multi-user requests.

Impact

• Improved Recall@5 by 9.4 percentage points and Precision@5 by 2.0 percentage points, while increasing Hit@5 from 84.0% to 100.0%.
• Reduced prompt memory tokens by 61.9%, from 737 to 281, while preserving multi-turn conversational context.
• Achieved 4.5× higher throughput under 10 concurrent multi-user agent requests in local benchmarks.

Metrics

Retrieval pipeline built over 72,877 chunks from 11,274 XML files and 16,407 question-answer documents; throughput measured locally under 10 concurrent multi-user requests.

Recall@5 improvement

+9.4 pp

Precision@5 improvement

+2.0 pp

Hit@5, up from 84.0%

100.0%

Throughput at 10 concurrent requests

4.5×

Technical Implementation

• Built a Redis Cache-Aside pipeline with cached null values, logical expiration, Redisson distributed locks, and asynchronous cache rebuilding to prevent cache penetration and breakdown.
• Built a two-level caching layer with Caffeine and Redis, using a Redis Lua sliding window to dynamically detect hot keys and promote frequently accessed data into local memory.
• Developed a Redis Lua flash-sale pipeline that atomically validates purchase eligibility, reserves inventory, and enforces one order per user before asynchronously creating orders through Redis Streams.
• Designed a reliable Redis Stream consumer workflow with consumer groups, idempotent processing, acknowledgments, pending-message recovery, retries, and dead-letter handling.
• Implemented a Spring scheduled order lifecycle workflow to batch-close expired unpaid orders, restore inventory, and resolve payment-cancellation races through compare-and-set status transitions.

Impact

• Eliminated database lookups and reduced P99 latency by 30.9%, from 43.88 ms to 30.31 ms, across 10,000 warmed-up queries under 100 concurrent requests.
• Achieved a 100% cache hit rate, increased throughput by 30.4% from 6,313.59 to 8,232.77 QPS, and eliminated Redis lookups in a warmed-up single-hot-key benchmark.
• Reached 3,668.06 QPS with 1,000 unique users, 100 concurrent requests, and 200 vouchers without overselling or duplicate orders.
• Reduced flash-sale API P99 latency by 83.2%, from 408.81 ms to 68.59 ms, while processing 309.15 orders per second.
• Produced zero invalid order states across 100 concurrency tests and closed expired orders within 60 seconds.

Metrics

Cache results measured with warmed-up queries under 100 concurrent requests; flash-sale results measured with 1,000 unique users and 200 available vouchers.

QPS with two-level caching

8,232.77

Flash-sale API QPS

3,668.06

Lower flash-sale P99 latency

83.2%

Oversold or duplicate orders

0