IRIS

IRIS (Integrated Revenue Intelligence System) is an AI-powered sales operations platform I built at Caylent. It lived directly in Slack, integrated with Salesforce, Gong, Jira, and Slack, and processed thousands of events daily to coordinate a rapidly growing sales team.

Why I built it

Sales operations at a growing company involve a lot of manual tracking, cross-system lookups, and process enforcement that nobody wants to do. Teams lose hours searching for call transcripts, updating CRM records, and chasing status on proposals and document approvals. Important things fall through the cracks because the information lives in the wrong system or nobody's watching.

IRIS solved this by being the single interface to the entire sales stack. Instead of switching between Salesforce, Gong, Jira, and Slack, the team talked to IRIS and IRIS figured out where to get the answer. It also handled things that were historically hard to track, like proposal review status, which didn't map cleanly to Salesforce and wasn't getting tracked at all before IRIS existed.

How it worked

The conversational layer used dual RAG systems built on AWS Bedrock. One handled structured data from Salesforce and DynamoDB. The other handled unstructured data like Gong call transcripts. The RAG engine generated DynamoDB queries from natural language, introspected table schemas to validate them, and when a query failed, it captured the error and retried with that context.

The agentic routing was the part I'm most proud of. IRIS determined which guided command or data source to use based on the user's natural language input, then orchestrated multi-step workflows with persistent conversation state. Commands were composable. Contexts were built programmatically from reusable prompt objects. The whole thing was meta-programmed to be extensible without touching core logic.

Conversations were structured as a tree. When a user activated a tool, it got its own child conversation. When the tool finished, that conversation was archived and a summary tombstone was placed in the parent. The AI kept context for what happened without retaining unnecessary details. This let conversations run much longer with low hallucination rates, because the model always had the right level of context without being overloaded with raw history.

IRIS also built team culture. When a deal closed, she posted to the wins channel with details, recognition for the pursuit team, and emoji tiering based on deal size. Big deals got a hype GIF first. It was a small feature that had outsized impact on morale and visibility.

She had a personality too. Old school hip hop vibes. Snoop, Dre, Cube. Pure engineering with character.