Inspiration In the modern enterprise, vendor compliance is a massive bottleneck where speed goes to die. Most companies spend weeks manually auditing vendor documents against complex internal standards. I wanted to build more than just a chatbot; I inspired to create an Active Agent that could live where teams already work—Slack—and take the heavy lifting out of risk assessment.
What it does VendorPulse is an automated compliance auditor that monitors Slack for vendor inquiries. When a user mentions the bot, it instantly triggers a multi-model workflow to audit the vendor against a 50-page compliance knowledge base. It identifies risks, categorizes vendor stability, and provides a professional executive-ready report directly in the Slack thread.
How we built it VendorPulse is powered by a multi-model orchestration workflow built on Airia:
The Trigger: A Slack Event Subscription monitors for @VendorPulse mentions, feeding the request into our Airia webhook.
The Auditor: Gemini 2.0 Flash acts as the primary investigator, cross-referencing vendor data against internal standards.
The Architect: Claude Haiku takes the raw audit findings and formats them into a structured Slack report.
The Human Gate: We integrated a Human-in-the-Loop approval block to ensure high-risk vendors are never cleared without a manual double-check from a compliance officer.
Challenges we ran into The biggest hurdle was the "Ghost Response" issue, where Slack would show a reply count but the thread remained empty. We had to debug the Interactivity Request URL and ensure the Slack Scopes (like chat:write) were perfectly aligned with Airia's webhook delivery. Overcoming the mismatch between Socket Mode and Webhook Events was the turning point that finally got the agent responding in real-time.
Accomplishments that we're proud of We are incredibly proud of successfully wiring a multi-model "relay" system. Getting Gemini and Claude to work together within a single Airia workflow to produce a real-world business output feels like a major technical achievement. Seeing the Human-in-the-Loop logic pause the agent's execution until a manual approval was granted proved that we built a truly "Active" and safe enterprise tool.
What we learned This project was a deep dive into asynchronous agent design. We learned that building a production-ready agent is as much about the infrastructure—managing Slack Interactivity, Request URLs, and Webhook Handshakes—as it is about the AI prompts themselves. We also gained experience in selecting specific models, like Gemini 2.0 Flash, for high-speed analysis.
What's next for VendorPulse We plan to expand VendorPulse to support automated document parsing. Instead of just checking vendor names, the agent will soon be able to scan PDF contracts and insurance certificates for "hidden" risk clauses using vision-language models. This will move compliance from a "human-assisted" process to a truly "zero-touch" automated workflow.
Built With
- claude-haiku
- gemini-2.0-flash
- iria
- slackapi
Log in or sign up for Devpost to join the conversation.