Projects
One live AI agent product, and four platform systems where the interesting part is what was traded away.
Family Learning Assistant
An AI agent product for families: a parent describes the goal in natural language, and the agent understands the intent, selects tools and orders the workflow — parsing study material, organising vocabularies, generating practice and managing learning tasks.
The core trade-offs: the agent never touches the database; the service layer is the only entry point for business rules, permissions, transactions and state checks; tool calling chooses which capability, workflow decides in what order; every create, update and delete is shown to the user before it runs.
Representative platform work
General terms, no confidential detailLBS platform
Location data and services for Ping An Bank across credit card and auto finance, from pre-loan to post-loan. Split into data services, high-frequency APIs, page services, admin and pre-production deployments.
External vendor gateway
One controlled channel between the bank and its vendors: routing, rate limiting, circuit breaking and caching, replacing vendor calls scattered across business systems.
Risk-control feature platform
Feature computation and variable accumulation for financial risk control, with a decision engine, case management and knowledge graph around it. I led replay, backfill and high availability, and delivered graph features, ClickHouse integration and sub-tenant design.
Marketing platform (PMS)
Tagging, event centre, placements, A/B testing and multi-channel reach. Offline and real-time tags share one definition layer that the engine resolves into different implementations; the event centre is configured through input, output and processing rules.
Described in general terms: no client names, internal addresses, trade secrets or unauthorised code. Technical detail can be discussed separately.