SmartAdmin
SmartAdmin is a capstone project that combines retrieval-augmented generation, large language models and browser-based automation to safely execute delivery modification workflows in uParcel’s logistics environment, with human-in-the-loop control and full auditability.
Walkthrough of the SmartAdmin architecture, RAG–LLM interpretation flow, approval dashboard and Playwright-based automation on the uParcel staging portal.
uParcel’s delivery modification tasks such as address changes, remark updates and urgency adjustments are currently handled through manual interaction with browser-based admin portals, which is time-consuming and error-prone when volumes scale.
SmartAdmin investigates whether an instruction-centric architecture using retrieval-augmented generation and large language models can interpret natural-language modification requests and translate them into structured action plans suitable for safe automation.
The system captures agent requests, retrieves relevant SOP fragments from a LanceDB-based knowledge base, and uses an LLM to generate JSON action schemas that describe the intended changes at a structured level. These actions are validated, gated by role-based policies and then executed via Playwright scripts on uParcel’s staging admin portal.
The capstone is structured around three core research questions that guide the design and evaluation of SmartAdmin.