Portfolio case study
An end-to-end demo for NordFit Store: an AI advisor that helps users choose fitness equipment based on goal, space, level and budget. The interaction can happen by text or voice, then high-intent conversations become CRM leads.
Real problem
In a technical catalog, visitors compare similar products, have space or budget constraints and often abandon before checkout. A generic FAQ can answer policies, but it does not guide the purchase or leave sales teams with useful follow-up context.
Solution
The NordFit Store demo shows a widget embedded in the shop. The assistant uses catalog and policy data, profiles the user and proposes a coherent setup instead of only retrieving static information.
The frontend extracts goal, space, level and budget from written or spoken messages. The logic scores products and suggests alternatives or accessories without silently exceeding budget.
n8n builds the prompt with knowledge base, chat history, current product and lead data. The required output is structured JSON with intent, product, setup, lead quality and confidence.
When purchase intent is high, the assistant asks for consent and minimum contact data. The workflow prepares Contact, Deal, Task and Note with conversation summary, estimated value and recommended setup.
Architecture
Static catalog, demo cart, product pages, quick view and embedded chat widget.
JavaScript detects preferences, ranks products, handles local fallback, voice input and spoken replies.
The webhook normalizes payloads, loads catalog context, sends the AI prompt and validates JSON output.
HubSpot receives qualified leads; email, logs and manual review handle outcomes and failures.
Proof of work
The value is not only the e-commerce UI. The project shows how to connect a consultative conversation to a controlled CRM process, with clear triggers and documented edge cases.
The advisor does not choose randomly: it compares goals, available space, level, budget and current product.
When useful, it suggests combinations such as bike plus mat or dumbbells plus bench, calculating the estimated total.
If the webhook does not respond, the widget continues with local logic instead of breaking the experience.
Email and phone are requested only after consent and only when real commercial intent emerges.
HubSpot is triggered only with consent, valid email, quote_request or ready_to_buy intent, confidence at least 0.70 and no manual review.
The team receives recommended product, estimated value, urgency, lead quality and conversation summary.
Beyond the demo
Questions about shipping, returns or warranty are answered from policies without creating CRM opportunities.
The system recommends bikes, rowers, treadmills, dumbbells or accessories based on constraints instead of pushing one item.
If the user leaves an email without consent, the workflow does not create HubSpot records and asks for explicit confirmation.
Messages asking to ignore rules, invent discounts or change policies are treated as untrusted input.
Stock, promotions and availability not present in the KB are not invented; the assistant offers sales follow-up.
Rate limits or unavailable APIs must produce logs, retry or manual review without losing the conversation.
Stack
The stack stays intentionally lightweight: static frontend for the demo, n8n as orchestrator, OpenAI for response and qualification, HubSpot for the sales pipeline.
HTML, CSS, vanilla JavaScript, product catalog, demo cart, chat widget, browser-native voice input.
Product scoring, preference extraction, quick prompts, recommended setup, local fallback if the webhook fails.
n8n, OpenAI, structured JSON output, catalog and policies as knowledge base, validation and manual review.
HubSpot Contact, Deal, Task, Note, estimated deal value, follow-up email with Resend or SMTP.
Need a similar system?
The demo uses fitness equipment, but the same pattern works in sectors where customers need to choose between technical products, configurations, budgets and sales follow-up.
Prefer email? Reach me at luigi.scorzelli87@gmail.com