WeBuildCrew
🤖 AI · Customer Support

Synapse — AI Support Assistant

An AI support assistant that deflects routine questions 24/7 by answering from the client's own documentation, with clean human handoff and full logging.

Timeline
5 weeks
Role
AI integration & full-stack
Team
Lead engineer
Year
2025
Next.jsTypeScriptOpenAIPostgreSQLpgvectorTailwind CSS
🤖
Synapse

62%

Tickets deflected

24/7

Coverage

<2s

Avg. response

0

Hallucinated policies

Client goals

  • Answer common questions instantly from real docs
  • Never invent policies; hand off to a human cleanly
  • Give the team visibility into every answer

Project overview

Synapse is an AI assistant that answers customer questions using the client's own documentation, not the model's guesswork. It runs 24/7, escalates cleanly, and logs everything for review.

The client problem

Our solution

We indexed the client's docs into a vector store, then grounded every answer with retrieval and a strict system prompt that defers to a human when confidence is low. The team sees every exchange.

What we built

Features

🧠

RAG grounding

Answers retrieved from the client's own content.

💬

Chat widget

Embeddable, on-brand chat for any site.

🙋

Human handoff

Escalates to support when unsure.

📚

Knowledge sync

Auto-indexes docs, FAQs and help articles.

🗒️

Full logging

Every Q&A logged and reviewable.

🔒

Guardrails

Strict prompts keep it on-topic and safe.

Under the hood

Technical code

app/api/chat/route.ts
TypeScript
export async function POST(req: Request) {  const { question } = await req.json();  const context = await searchKnowledgeBase(question); // RAG grounding  const answer = await ai.chat({    system: "Answer ONLY from the provided context. Offer a human if unsure.",    messages: [      { role: "system", content: context },      { role: "user", content: question },    ],  });  return Response.json({ answer });}
app/api/projects/route.ts
TypeScript
import { db } from "@/lib/db"; export async function GET() {  const items = await db.project.findMany({    where: { status: "active" },    select: { id: true, name: true, updatedAt: true },    orderBy: { updatedAt: "desc" },    take: 50,  });  return Response.json(items);}

Integrations

APIs & services

OpenAI

LLM responses with strict system prompts.

pgvector

Vector search over the knowledge base.

Resend

Escalation emails to the support team.

Slack

Live handoff notifications.

Process

Development workflow

  1. 1

    Discovery

    Goals, users and success metrics defined on a free call.

  2. 2

    Planning

    Fixed scope, milestone plan and timeline in writing.

  3. 3

    UI/UX design

    Wireframes and a clean, modern design system.

  4. 4

    Backend setup

    Database schema, auth and core services.

  5. 5

    API development

    Typed, documented endpoints and integrations.

  6. 6

    Frontend

    Responsive, accessible UI wired to the API.

  7. 7

    Testing

    QA across devices plus performance & security checks.

  8. 8

    Deployment

    CI/CD, monitoring and a smooth production launch.

  9. 9

    Support

    Ongoing maintenance, updates and new features.

Speed & SEO

Performance & SEO

Performance

  • Server-side & static rendering for instant first paint
  • Image optimisation and lazy loading
  • Code-splitting and route-level prefetching
  • Edge/CDN caching with smart revalidation
  • Green Core Web Vitals (LCP, CLS, INP)
  • Database indexing and query tuning

SEO

  • Per-page metadata, canonical URLs and Open Graph
  • Structured data (JSON-LD) for rich results
  • Programmatic sitemap & robots
  • Semantic, accessible HTML
  • Clean, descriptive slug URLs

Built secure

Security

🔒Session-based auth with role permissions
🔒Server-side authorization on every protected route
🔒Input validation (Zod) on client and server
🔒Secure payments via hosted Checkout (no card data stored)
🔒Rate limiting and abuse protection
🔒Encrypted secrets and least-privilege access

Outcome

Results & impact

62%

Tickets deflected

24/7

Coverage

<2s

Avg. response

0

Hallucinated policies

Synapse deflected 62% of routine tickets within weeks, responded in under two seconds, and produced zero hallucinated policies thanks to retrieval grounding and guardrails.