I build systems that survive real delivery pressure.

Delivered a production AI guide service with POI knowledge retrieval, route/weather tools, multi-turn memory, and fallback handling for weak-network scenarios.

• Built RAG pipeline for scenic spots, routes, and check-in points with chunking, embeddings, layered retrieval, and answer-quality evaluation.
• Integrated Spring AI and LangChain4j for function calling across POI search, route planning, weather, and safe fallback responses.
• Implemented session-based memory with sliding-window context so location, preference, and itinerary survive multi-turn conversations.

Built and maintained a multi-source map-road data platform handling trajectory, UGC, simulation, and toolchain inputs for downstream traffic and topology workflows.

• Used Conductor workflows and DMQ async listeners to process million-record ingestion jobs with clearer observability.
• Designed a Python SPI quality-check mechanism so data rules can be updated without rewriting the Java platform.
• Optimized batch JDBC and PostgreSQL indexes, moving core processing from days to hours and improving query performance by 30%+.

Standardized acquisition, normalization, diffing, cleaning, aggregation, and publishing for transit data across China, Japan, and Korea.

• Supported full transit-data replacement for 371 Chinese cities and overseas metro coverage for Japan and Korea.
• Implemented stop-to-station fusion using multi-source cross-validation to improve pickup-point accuracy.
• Built scheduled automation with phased controls and integrity checks; client reporting showed 97.5% station-cleaning automation and 98% accuracy.

Built automated tile-data ingestion and offline workbench features for pedestrian facility data.

• Implemented phased XXL-JOB workflows with data-integrity checks for automated acquisition, cleaning, and database output.
• Delivered offline editing, road-network consistency checks, state transitions, and export/publishing modules.
• Supported pedestrian-facility coverage across 30 Indian cities and a client-reported 3% road-network consistency improvement.