Support knowledge was scattered
Useful answers lived across file formats, docs, and product references rather than one customer-queryable surface.
FIELD NOTES / B2B SAAS & CUSTOMER SUPPORT AI
Customer support.
Grounded before it replies.
Suprix turns every customer organization’s documents into its own isolated AI support brain, then connects that knowledge to an embeddable website widget, voice support, and a human operator workflow when automation should step back.
AI SUPPORT CASE STUDY
Suprix
Multi-tenant RAG.
Widget and voice support.
Suprix / Operator command centerGROUNDED SUPPORT EXPERIENCE
AI + OPERATOR LOOPAutomation answers where knowledge exists.
When coverage is thin, the conversation keeps its messages, customer context, and lifecycle state for a human operator.
Follow the handoff ↗IngestRetrieveAnswerEscalateResolve
DOMAIN
B2B customer support
WHAT WE BUILT
Multi-tenant AI support platform
OUR ROLE
RAG, product & full-stack engineering
01 / THE SUPPORT PROBLEM
The answer exists.
The customer still waits.
Most support teams already have the answer somewhere: product docs, PDFs, policies, pricing sheets, onboarding notes, SDK guides, images, or previous support material.
The hard part is making that knowledge safely available at the exact moment a customer asks—without letting AI invent an answer when the organization’s own material is silent.
Suprix solves the problem as an operating system for support automation: tenant-isolated knowledge, grounded responses, real-time conversations, and human escalation inside one workflow.
Generic chatbots were risky
Pricing, policy, deployment, and support answers needed to come from approved organization material.
Escalation had to preserve context
Operators needed the full conversation, customer profile, device data, language, and unresolved status.
02 / INSIDE THE PRODUCT
AI where it fits.
Humans where it matters.
The product combines an authenticated operator dashboard, a knowledge-management surface, an embeddable widget, and optional voice support—so customer interactions do not fragment across separate tools.
AI answers first, operators keep the final say.
The inbox shows AI-assisted conversations beside customer context, status, and escalation controls. Operators can reply, enhance drafts, and resolve only when the interaction is genuinely handled.
CUSTOMER SUPPORT LOOP
Website visitor→Widget session→Organization RAG search→Grounded reply→Escalation if needed→Operator resolution
03 / THE PLATFORM DESIGN
Tenant isolation is not an afterthought.
It is the retrieval boundary.
Suprix isolates knowledge at the organization namespace level before retrieval happens. That means the AI does not search a global pile of passages and filter later—it starts inside the right organization’s approved material.
Look under the hood ↘01
Ingest organization knowledge
Admins upload support docs, SDK material, product catalogs, policy files, images, HTML, and text content.
Embedded widget
Visitors ask questions through a lightweight chat surface installed on the business website.
Human dashboard
Support teams review unresolved conversations, customer context, and AI-generated drafts.
02
Answer only from the right knowledge
The AI retrieves within the customer organization’s namespace and uses that context to generate a grounded reply.
03
Escalate instead of guessing
If coverage is missing or the customer asks for a person, the entire interaction moves into operator workflow.
One support journey, several controlled surfaces.
Every surface uses the same conversation record, organization scope, and customer context.
Docs→Knowledge namespace→Widget→AI reply↘Operator handoff
Support automation—not unattended authority.
Suprix is designed to answer documented questions, capture context, and escalate unresolved cases. The product value comes from combining AI speed with operator governance.
04 / THE TECHNOLOGY & THE REASON
A modern support stack with
clear separation between tenants, channels, and operators.
The implementation keeps product state, identity, knowledge retrieval, widget delivery, voice integration, and human workflow as separate concerns—so the platform can scale without mixing customer data or support responsibilities.
WEB APPLICATION
Next.js + React
Powers the operator dashboard, configuration screens, and customer-facing widget experience.
Why it matters
One component system can serve admin, support, and embedded surfaces while keeping each experience tailored.
REAL-TIME BACKEND
Convex
Stores conversations, messages, widget configuration, organization-scoped data, and real-time updates.
Why real-time state?
The customer widget and operator dashboard need the same conversation state without fragile polling.
RAG + GENERATION
OpenAI models
Supports retrieval-aware response generation and operator draft enhancement.
Why grounded generation?
The model’s role is to compose from approved retrieved context, not replace the organization’s source of truth.
OPERATOR IDENTITY
Clerk
Separates authenticated organization members from anonymous or lightweight customer sessions.
Why two identity models?
Operators need authenticated access; website visitors need low-friction support entry.
VOICE CHANNEL
Vapi integration
Adds AI voice-call capability as a configurable support channel.
Why voice belongs here
Voice support becomes another governed support path rather than a separate, untracked assistant.
SECRETS + CONFIG
AWS Secrets Manager
Keeps third-party credentials and integration configuration outside application code.
Why isolate secrets?
Support channels, AI providers, and voice integrations need managed credentials with clear operational boundaries.
Upload knowledge
Admins add organization-owned documents, catalogs, HTML, images, PDFs, and text.
Create tenant scope
Content is prepared for retrieval inside the owning organization’s namespace.
Load the widget
A small script initializes the support interface on the customer’s website.
Search before replying
The AI uses retrieved passages to produce customer-facing support answers.
Escalate cleanly
Operators receive unresolved cases with customer context and conversation history.
The supplied screenshots show the operator inbox, knowledge base, Vapi configuration, and widget states. Production metrics depend on each customer’s documents, usage volume, escalation policy, and quality review process.
Explore the documented platform stack
- Frontend
- Next.js 15 · React 19 · TypeScript · embeddable widget · iframe-isolated customer surface
- Backend
- Convex real-time data · conversations · messages · customers · widget configuration · organization state
- AI layer
- OpenAI GPT-4o · retrieval-augmented generation · operator draft enhancement · fallback behavior
- Identity
- Clerk for operators and admins · lightweight customer sessions for website visitors
- Integrations
- Vapi voice calls · billing infrastructure · third-party support configuration
- Operations
- Secrets management · file storage · scheduled operations · webhook processing
Support boundary
Suprix should not invent policy, pricing, procurement, or product answers when organization knowledge is missing. The safe behavior is to ask for clarification or move the conversation to a human operator.
Technology names identify the documented implementation, not partnerships or endorsements.
05 / WHY THIS MATTERS
Customers get speed.
Teams keep control.
The strongest part of Suprix is not the chat bubble. It is the operating model behind it: organization-owned knowledge, controlled retrieval, real-time conversation state, and a fallback path that respects the limits of automation.
That makes the platform valuable for SaaS teams, product-led companies, procurement-heavy support desks, and any business where customer answers must match the current approved material.
KNOWLEDGE
Multi-format ingestion · organization-scoped retrieval
CHANNELS
Website widget · chat · optional AI voice calls
OPERATIONS
Operator inbox · context panel · lifecycle status
CONTROL
Fallback · escalation · human resolution
AI starts with search
Responses are grounded in organization material before they are customer-facing.
One tenant never trains another
Knowledge access is scoped to the owning organization’s namespace and configuration.
Operators inherit the whole story
Escalation includes the conversation, customer identity, device context, and unresolved status.
The widget is easy to adopt
A simple script-based install keeps deployment friction low for customer websites.
A CONVERSATION ABOUT YOUR SUPPORT WORKFLOW
Where should AI answer,
and where should it hand off?
Bring one support workflow, product catalog, or customer-facing knowledge base. In 20 minutes, we’ll map the tenant boundary, retrieval policy, widget experience, and escalation design.
20 minutes · Your support workflow, constraints, and next step
20 minutes · Your support workflow, constraints, and next step