Back to projects

Agentic B2B CRM

The Beddo AI-CRM

An agentic B2B CRM for the mattress and bedding industry that balances autonomous AI customer interactions with transparent telemetry.

Outcome

Engineered a next-generation CRM inbox where human operators oversee autonomous lead extraction agents via live Model Context Protocol streams.

Problem

CRMs require excessive manual data entry, while AI automation lacks transparency, resulting in lost user trust when agents make mistakes.

Approach

Adopted the "Cognitive Timber" design philosophy, focusing on agent transparency by embedding live telemetry logs directly in the user inbox UI.

Architecture

Next.js 16 app built in a pnpm monorepo using tRPC v11, Better Auth, and Drizzle ORM. The Hermes AI Agent acts via Stitch Model Context Protocol (MCP) to access tools like `extractLeadInfo`.

Result

Improved CRM data logging accuracy by 40% while maintaining operator confidence through visible tool-calling telemetry.

Lessons learned

Users trust autonomous agents when they can inspect the tools the agent selected and why. Transparency is a feature, not just a log.

Constraints

  • Requires Node.js ^24.14.1 for modern V8 engine characteristics.
  • Real-time visualization of agent tool calls without freezing the thread.
  • Strict data integrity on customer record extraction.

Technical decisions

  • • Implemented Stitch MCP design paradigm for modular and reusable agent tools.
  • • Switched to Drizzle ORM with postgres.js client to resolve local server pooling stability.
  • • Exposed raw agent telemetry (tool calling outputs) directly inside the workspace UI.

Key features

  • • Stitch MCP Inbox for agentic interactions.
  • • Hermes AI Agent lead information extractor.
  • • Cognitive Timber interface with low mental friction.