What Inspired Us
In global logistics, every shipment involves complex contracts, variable rate agreements, and hundreds of line items in invoices. We noticed that even small mismatches — like overbilling, duplicate entries, or excessive surcharges — can silently leak revenue across thousands of transactions. Inspired by real-world challenges faced by logistics, finance, and supply chain audit teams, we set out to build an AI-first, multi-agent system that could autonomously detect such anomalies and explain them with human-like clarity.
What we learned
Through this hackathon, we deep-dived into Google’s Agent Development Kit (ADK) and learned how to build modular, LLM-powered agents that could collaborate and share context through structured session memory. We gained experience orchestrating agents with SequentialAgent, validating inputs using Pydantic-based schemas, and implementing interactive, production-style Streamlit UIs. We also learned how to blend traditional business logic (e.g., billing rate lookups) with generative AI for flexible natural language interaction and automation.
How we built it
- We started with structured contract PDFs and invoice Excel sheets, extracting both tabular and textual data using ContractExtractorAgent and InvoiceExtractorAgent. 2.Then, we developed an AnomalyDetectorAgent to classify user queries into anomaly types and filter relevant data using logic-based detection rules. 3.An AnomalyQAAgent further explained anomalies in conversational form, and an EmailAgent was designed to send automated reports to vendors when issues were found. 4.We integrated all agents into a SequentialAgent pipeline using the ADK’s Runner, enabling a complete end-to-end flow — from file upload to anomaly detection to explanation and email drafting.
- A Streamlit-based UI served as the user-friendly frontend for uploading files and asking follow-up questions.
Challenges we faced
1.Strict schema validation in ADK (e.g., Session creation, run_async input format) required deep debugging and careful structuring of agent inputs and outputs. 2.Achieving asynchronous agent execution and maintaining chat continuity in Streamlit without infinite loops took multiple rounds of session state management fixes. 3.Designing detection logic that balanced hard-coded business rules with flexible AI-driven classification was both a modeling and engineering challenge. 4.Most importantly, we had to build for scalability and automation, ensuring that agents worked even in the absence of user queries — pushing us to create self-running detection flows.
Log in or sign up for Devpost to join the conversation.