Keystone CRM: AI-powered CRM with an agentic assistant, AI lead scoring and unified inbox

A UK B2B events business ran its sponsorship, exhibitor and speaker pipeline across spreadsheets, WhatsApp threads and an older PHP CRM. Leads were tracked in sheets, conversations lived in personal inboxes and messengers, and follow-ups depended on memory. Deals slipped, nobody had a single view of each customer, and sales staff spent time reading long message threads and writing replies instead of selling. The business needed one system that captured every lead and conversation, prioritised who to contact, and let non-technical staff adapt the CRM themselves without a developer.

I designed and built Keystone CRM end to end (product design, architecture, database, AI integration, messaging integrations and migration). Core modules: leads, deals, quotes, tasks, notes, calendar and one inbox for email (IMAP), WhatsApp and SMS (Twilio). AI layer: every lead gets an AI score; inbound replies are tagged for mood and intent with an AI-drafted response; an AI assistant with 13 tools searches, creates and updates records, builds pages/columns/dashboard widgets and sends messages, by typed or spoken request, always asking for confirmation before acting, with every AI write or send logged in an audit trail. RAG: uploaded PDFs/Word documents and CRM data are chunked and embedded in PostgreSQL and retrieved (LangChain/LangGraph) for meeting briefs, answers and lead-data extraction from message threads. Models: Claude Haiku for fast text, Claude Sonnet for the agent, NVIDIA Nemotron fallback with tool-call translation. Stack: Node.js/Express, PostgreSQL (JSONB custom-pages engine so teams add pages and columns without migrations), Docker. Includes white-label branding, role-based access, rate-limited webhooks and a side-by-side migration from the old PHP CRM.

In live production: the events business now runs its full sponsorship, exhibitor and speaker pipeline in one CRM instead of spreadsheets, WhatsApp and a legacy system. Every lead is AI-scored so staff know who to contact first, and every inbound reply arrives tagged with mood and intent plus a draft response, cutting the time spent reading threads and writing replies. Staff build their own pages, columns and dashboards (including dashboards generated from one sentence) without developer help. The migration from the old PHP CRM ran side by side with no re-entry of mail credentials and old tables were removed only after verification. The product is now offered to other businesses as white-label Keystone CRM (buy outright or subscribe). Product demo video (3 min): https://youtu.be/PGXzohwN26U