Portfolio case study
An end-to-end demo connecting a web chat widget, WhatsApp, a private knowledge base, n8n, pgvector and HubSpot CRM to answer patients and qualify appointment requests.
Real problem
A professional clinic receives requests about opening hours, prices, urgent cases, insurance and appointments. A generic chatbot is not enough: the system must connect to real channels, the clinic knowledge base and the CRM used by the team.
Solution
The Studio Dentistico Rossi landing shows the customer-facing side. Behind it, n8n orchestrates AI, a vector database, CRM updates and notifications to turn conversations and forms into operational work.
The patient writes from the website or WhatsApp. The workflow retrieves context from the knowledge base, keeps conversational memory and answers only from controlled information.
The form sends a structured payload to n8n. AI identifies urgency, requested service, estimated value and the next action for the front desk.
The workflow checks duplicates, creates or updates the contact, opens a deal, creates a task and stores a note with the AI summary of the request.
Architecture
WhatsApp Business Cloud API, chat widget and appointment form on the demo landing.
n8n receives webhooks, normalizes data, handles routing, errors and frontend responses.
OpenAI creates embeddings, searches the knowledge base through pgvector and returns contextual answers.
Neon PostgreSQL stores memory and documents; HubSpot receives leads, deals, tasks and operational notes.
Proof of work
The important part is not the clinic landing, but the system behind it: real channels, controlled knowledge, CRM automations and predictable behavior when something goes wrong.
Services, opening hours, prices, urgent cases, insurance, medical team and booking rules are collected in a KB file, split into 1000-character chunks and retrieved with cosine top-4 search on pgvector.
The chat uses a session ID or phone number as the key and keeps a 10-message window, so each turn stays consistent with the previous exchange instead of being isolated.
The widget can return demo answers even when the webhook is unavailable, avoiding a broken UX during tests or maintenance.
The lead does not remain in an email: the AI returns urgency, service and estimated value as structured JSON, which becomes a HubSpot contact, a deal in the pipeline, a task within 24h and a note for the front desk.
Out-of-scope, urgent or ambiguous requests can be logged and passed to a person instead of forcing an AI answer.
The repo documents credentials outside the frontend, fictional data, tokens in n8n credentials and no real health data in the demo.
Beyond the demo
The answer must stay grounded in the KB and suggest human contact when the request cannot be handled by the assistant.
Pain, swelling, trauma or broken teeth must trigger high priority and fast-contact guidance, without medical diagnosis.
Messages trying to change operating rules must not alter amounts, priorities or workflow behavior.
Timeouts, unavailable webhooks, HubSpot 429s or failed email should produce logs, retries or manual review without losing the lead.
Stack
I chose cost-conscious, realistic tools for small businesses: static hosting for the landing, n8n for orchestration, Postgres with pgvector for RAG and HubSpot as the sales CRM.
Astro, React, Tailwind CSS, chat widget, appointment form, Vercel deployment.
OpenAI, GPT-4o-mini, text-embedding-3-small, retrieval over a private knowledge base.
Neon PostgreSQL, pgvector, document table, chat memory, messages and handoff requests.
n8n, webhooks, WhatsApp Business Cloud API, HubSpot CRM API, SMTP or Resend.
Need a similar system?
The demo is built around a fictional dental clinic, but the same pattern works for clinics, consultants, hospitality, education, real estate and professional services.
Prefer email? Reach me at luigi.scorzelli87@gmail.com