Technical Architecture

Multi-agent orchestration with circuit breakers and structured output across 5 LLM providers

11-node LangGraph intent classifier routing 3 paths: hybrid RAG QA, CatBoost+SHAP cost ML, and sandboxed code-gen data analysis

LLM-as-Judge evaluation pipeline with drift detection and incident triage workflow

Local LLM decision engine with 7-trigger hybrid entry and 6-layer exit management

Dual ASR engine pipeline with AI correction and pronoun verification across 3 interfaces

Field-level AES-256-GCM encryption with macOS Keychain integration and Unix socket API

Plugin architecture player with dual API signing and region unlock proxy system

Enterprise flash-sale engine with zero overselling under 10K+ concurrent buyers — Redis atomic stock deduction + Lua, RocketMQ async order pipeline, layered defense-in-depth, and a self-healing compensation loop that reconciles stuck inventory

Company-wide authorization platform for ~1,000 employees across a dozen systems — menu/button/URI 3-tier RBAC, Redis-first hot path, and binlog CDC → RocketMQ cache sync

Deposit-free leasing credit risk engine spanning pre/in/post-lease lifecycle — multi-bureau scoring (A/B/C/F/L/O cards), 4-tier approval, and dynamic credit-limit pools

Ticket-purchase Saga across four microservices — Temporal orchestration, transactional outbox → RabbitMQ → live SSE, an idempotency barrier table, and automatic compensation in reverse on failure

A thin MCP bridge over a traditional Go ticketing backend — LLM agents act only through tools/resources/prompts behind a human consent + scope gate, with concurrency-safe booking and messaging off the critical path

Six-source financial ETL for HKEX IPO compliance — RocketMQ choreography gated by an Oracle master-table state machine, request_id idempotency, dynamic-datasource @DS pool isolation, and Fork/Join divide-and-conquer extraction staged in Redis; consistency without a distributed lock

Offline scorecard-modeling pipeline — WOE/IV feature engineering, logistic-regression scorecards with GBDT/anomaly/graph/survival challengers, KS/PSI validation — writing weight-optimization advice back to the online credit engine; model choice driven by data characteristics, not the newest algorithm