AI Sales Agent System for HubSpot — Lead Qualification, Scoring & Automated Follow-Up

I have a B2B lead-generation agency working with SMB clients across logistics and professional services. Every day we get 80–150 inbound leads through web forms, LinkedIn outreach replies, and email, all landing in HubSpot. Right now, a human person manually reads each one, decides if it's qualified, scores it, and figures out who follows up and how. It's slow (leads often sit 4 to 6 hours before first contact) and inconsistent: different reps qualify leads differently. I want an AI system that sits on top of our HubSpot CRM and: 1. Reads every new lead the moment it comes in 2. Uses an LLM-based agent to understand context: industry, company size, message intent, urgency 3. Scores the lead using a model trained on our historical won/lost deal data 4. Automatically drafts a personalized first-touch email reply 5. Escalates high-value or ambiguous leads to a human person instantly, and routes low-priority leads into a nurture sequence 6. Logs every decision back into HubSpot so we keep a full audit trail We have about 14 months of historical CRM data (won/lost outcomes, deal size, response times) we can hand off for the scoring model. What I Need * An agentic workflow (LangGraph or similar) that owns the lead-triage decision end to end, not just a single prompt call * A RAG layer over our sales playbook and past won-deal transcripts, so the agent's replies sound like us, not generic AI * A lightweight ML lead-scoring model trained on our historical HubSpot data * Native HubSpot integration via API + webhooks: real-time, not batch * n8n (or equivalent) orchestration for the parts that don't need to live inside the agent itself * A real deployed service, not a local script — Dockerized, hosted, with basic uptime monitoring, since this runs 24/7 in production * A simple internal dashboard or Slack alert for escalation cases Deliverables * Working agentic pipeline connected live to HubSpot * Lead-scoring model + a short evaluation write-up (accuracy/precision on held-out historical data) * n8n workflow exports * Deployed back-end (Docker + cloud host) with monitoring * Documentation + a short walkthrough for our ops team Tech Stack I am open to: Python · LangGraph / CrewAI · OpenAI or Anthropic API · FastAPI · Docker · HubSpot API · n8n · AWS (or similar) Please ask me clarifying questions