How it works · Lead Assistant

The flow

One enquiry, from first message to founder review

01
Visitor
Asks a question in the website chat widget.
02
Safety screen
Pattern checks and Meta Prompt Guard stop prompt injection, abuse and spam before the agent runs.
03
AI assistant
A Groq-hosted open-weight model (gpt-oss-120b) answers from approved business information and calls tools.
04
Qualification
Code applies the business rules (service, budget, timeline, details). The model never decides the outcome.
05
Lead management
Details, events and decisions are saved in PostgreSQL and streamed to the live panel.
06
Founder review
The dashboard shows every lead; follow-up drafts wait for a person to approve them.

The only actions the model can take. None of them sends a message.

update_lead
Saves details the visitor shared, normalised (budget, timeline, email), then re-checks the rules.
evaluate_qualification
Re-applies the business rules to the saved lead.
get_open_slots
Lists open discovery-call times for a qualified lead.
book_slot
Books a time the visitor picked. Confirmed only after the booking is saved.
flag_for_founder
Hands the conversation to a person (human request, unanswerable question, data deletion).

Stated plainly, so you know exactly what you are testing

WorkingWebsite chat with streaming repliesVercel AI SDK UI stream from FastAPI
WorkingAI model and tool callingGroq free tier; a custom Python tool loop, no agent framework
WorkingBusiness-rule qualification and fit scoreRules in code; score from gpt-oss-20b
WorkingPrompt-injection and abuse screenPatterns + Llama Prompt Guard 2
WorkingLead database, events and dashboardPostgreSQL, viewer-scoped API
WorkingFollow-up drafts with human approvalDrafted by the model, decided by a person
WorkingAutonomy setting: book instantly or request approvalPer visitor session; outbound messages stay locked to approval
WorkingTalk-to-a-person hand-off with a briefEmail asked first; brief built from saved data
WorkingReply feedback, insights and coaching listFrom stored records only
Test modeDiscovery-call bookingLocal calendar from working hours; no invites sent
Not builtSending follow-up emailsNo email provider connected
Not builtWhatsApp, voice and CRM integrationsNot part of this version

Model reasoning, prompts and raw tool arguments never reach the browser; the activity feed uses fixed, plain labels.
Each visitor sees only example leads and their own; any other id returns “not found”.
Emails and phone numbers are masked before traces leave the server.
Public endpoints are rate-limited; visitor data is deleted after 7 days.