portfolio / 2026Coding with Jorlin
Back to work

03 / Lead decisions & customer follow-up

AI Sales Lead Decision & CRM Automation

Turning scattered customer facts into explainable next actions.

Enter system View GitHub

System preview

AI Sales Lead Decision & CRM Automation

A real capture of the current product interface. Select the preview to enter an available demo.

https://leadflow-sales-demo.onrender.com/AI Sales Lead Decision & CRM Automation System previewEnter system
Project overview

An end-to-end decision prototype for sales leads, scoring, customer facts, and CRM routing.

Who it is for

Sales operations teams handling leads from email, web forms, and other channels that need consistent decisions, replies, and CRM records.

What it does

  • Match products from manuals with evidence and fit constraints
  • Preserve customer preferences across turns and draft channel-aware replies
  • Support Qwen-grounded drafting, human review, and CRM / Notion synchronization

How to explore it

  1. Enter a request such as: Recommend a 7-day Yunnan trip for four people.
  2. Review the recommended product, fit reasons, and manual evidence.
  3. Add hotel, vehicle, and guide preferences in the same conversation and review the updated reply.
  4. Public visitors use an isolated demo; live models and Notion delivery require protected operator configuration.

The public demo uses offline manual retrieval and rule-based replies, with no paid LLM calls or real Notion writes. Free hosting may start slowly and demo records may reset after a restart. Do not enter real customer data.

Enter system

The challenge

Customer information is spread across text, forms, and history. A model suggestion without evidence is difficult to audit.

The approach

  • FastAPI / Pydantic manage typed facts and conversation state; Qwen supports understanding and grounded drafting, with deterministic product constraints.
  • BGE-M3, BM25, and RRF retrieval are supported; the public demo uses offline keyword-RRF and the same product manual snapshot.
  • Evidence, fit constraints, human review, PostgreSQL persistence, and retryable CRM delivery are supported; public visitors cannot perform real external writes.

The outcome

The existing 18-query retrieval evaluation reached 18 / 18 Hit@1 and Hit@3. The latest regression passed 699 tests with 95.21% coverage; two PostgreSQL tests were skipped without a local database.

Boundaries & reflections

The evaluation is intentionally small and validates feasibility; real deployment needs more industry data and review.