Multi-Agent AI Copilot for Small Business Admin — Customer Messages → Quotes, Invoices & Replies
While cold-pitching web development services to small business owners across Bengaluru — CCTV installers, electrical contractors, home-service providers — a pattern emerged immediately.
Their biggest daily friction wasn't marketing or accounting. It was the gap between a customer WhatsApp message and a professional reply.
A typical request arrives as:
"need quote for 3 cctv cameras with installation"
What follows takes 10–15 minutes of manual work:
- Recall or look up unit prices from a price list or memory
- Manually calculate quantities, subtotals, GST, and totals
- Type out a formatted quotation in WhatsApp or Word
- Copy-paste the reply back to the customer
This happens dozens of times a day. Pricing errors are common. Response times are slow. For a solo operator or a 2-person shop, this is lost revenue and wasted expertise.
BizPilot AI solves it end-to-end.
BizPilot AI converts raw natural language customer messages into professional, verified business outputs — quotations, invoices, policy replies, and follow-up messages — through a transparent 5-agent AI pipeline, in under a minute.
The business owner types (or pastes) the customer's message. BizPilot does the rest:
Customer WhatsApp message
↓
[1] Intake Agent — classify intent, extract entities
↓
[2] Context Agent — match products, look up pricing, check loyalty
↓
[3] Generation Agent — produce the document + human-readable reply
↓
[4] Approval Agent — auto-approve or route to owner for high-value orders
↓
[5] Review Agent — verify every price, line total, and GST figure
↓
Verified Quote / Invoice / Reply → one-tap WhatsApp share or PDF download
Every step is visible in the live pipeline view, so the owner stays in control — they see exactly what each agent decided and why.
- 5-agent transparent pipeline — Intake → Context → Generation → Approval → Review, each with a live status feed visible to the user
- Self-verifying output — the Review Agent cross-checks every unit price and line total against the master product catalog before the output is shown; if anything is wrong, it triggers one automatic regeneration cycle
- Returning-customer loyalty discount — the Context Agent tracks visit history in-session; repeat customers automatically receive a 5% loyalty discount, reflected in the document and GST calculation
- High-value order approval routing — orders above ₹12,000 are automatically flagged and routed to the owner (simulated Rahul Sharma approval) before the quote is finalised
- Multilingual replies — generated customer-facing messages can be switched between English, Kannada (ಕನ್ನಡ), and Hindi (हिंदी)
- One-tap WhatsApp sharing — the formatted reply opens directly in WhatsApp (
wa.medeep link) with the full message pre-filled; no copy-paste needed - Professional PDF export — quotes and invoices download as properly formatted A4 PDFs with line items, GST breakdown, and company letterhead via jsPDF
- Rule-based fallback — if the Gemini API is unavailable or rate-limited, every agent falls back to deterministic rule-based logic; the app stays fully functional
- Fully mobile-responsive — dedicated mobile layout with an accordion pipeline view and a bottom-sheet document preview
| Technology | Role |
|---|---|
| React 19 + Vite 8 | UI framework and dev server |
| Tailwind CSS 4 | Utility-first styling |
| Lucide React | Icon set |
| jsPDF | Client-side PDF generation |
wa.me deep links |
WhatsApp message pre-fill and share |
| Technology | Role |
|---|---|
| Node.js + Express 4 | HTTP server and pipeline orchestrator |
@google/generative-ai |
Gemini 2.5 Flash — intent classification, generation, review |
@anthropic-ai/sdk |
Claude — reserved for fallback classification |
uuid |
Unique request IDs for pipeline tracing |
dotenv |
Environment variable management |
| In-memory store | Customer visit tracking (loyalty discount state) |
| Component | Details |
|---|---|
| Intake Agent | Gemini 2.5 Flash with few-shot prompting; classifies into quote_request, invoice_request, customer_query, follow_up, unclear |
| Context Agent | Rule-based product matcher + fuzzy alias lookup against seed.json catalog; loyalty check via visit counter |
| Generation Agent | Gemini 2.5 Flash structured output; produces document JSON + human-readable multilingual reply |
| Approval Agent | Deterministic threshold logic (₹12,000); no LLM required |
| Review Agent | Gemini-assisted + deterministic pricing verifier; triggers one regeneration cycle on failure |
| Product Catalog | backend/data/seed.json — 12 products & services with aliases for fuzzy matching |
| Layer | Platform |
|---|---|
| Frontend | Vercel — biz-pilot.vercel.app |
| Backend | Render — Node.js web service |
🚀 https://biz-pilot.vercel.app
Try these demo inputs to see the full pipeline:
"Need quote for 3 CCTV cameras with installation"→ generates a full itemised quotation"Create invoice for 2 CCTV cameras and one DVR"→ generates a GST invoice with line totals"What is your refund policy?"→ pulls from the company policy document context"Follow up with customer who asked price yesterday"→ drafts a follow-up message
Note on response time: Live requests go through the Gemini 2.5 Flash API; first responses may take 3–8 seconds depending on API load. If the API is unreachable, the pipeline automatically switches to rule-based mode — all features remain functional, the only difference is the reply text is template-based rather than LLM-generated.
- Node.js ≥ 18
- npm ≥ 9
- A Google AI Studio API key (free tier works)
git clone https://github.com/Afnan-0206/BizPilot.git
cd BizPilotcd backend
npm installCreate your environment file:
cp .env.example .envOpen .env and fill in your Gemini API key:
GEMINI_API_KEY=your_actual_api_key_here
GEMINI_MODEL=gemini-2.5-flash
PORT=3001
NODE_ENV=developmentStart the backend:
npm run dev # development — auto-restarts on file changes (nodemon)
# or
npm start # productionThe backend will be live at http://localhost:3001. Verify with:
curl http://localhost:3001/healthOpen a new terminal window:
cd frontend
npm install
npm run devThe frontend will be live at http://localhost:5173.
From the project root, double-click start.bat or run it from a terminal:
start.batThis opens both the backend and frontend in separate terminal windows simultaneously.
| Variable | Required | Default | Description |
|---|---|---|---|
GEMINI_API_KEY |
Yes (for AI mode) | — | Google AI Studio API key |
GEMINI_MODEL |
No | gemini-2.5-flash |
Gemini model identifier |
PORT |
No | 3001 |
Backend server port |
NODE_ENV |
No | development |
Node environment |
No API key? The backend runs in rule-based mode automatically when
GEMINI_API_KEYis absent or set to the placeholder value. All 5 pipeline agents still execute — LLM-generated text is replaced with deterministic templates, but every other feature (pricing, PDF, WhatsApp, loyalty discounts, approval routing) works exactly the same.
bizpilot-ai/
├── backend/
│ ├── agents/
│ │ ├── intake.js # Agent 1 — intent classification & entity extraction
│ │ ├── context.js # Agent 2 — product matching, pricing, loyalty
│ │ ├── generation.js # Agent 3 — document + reply generation
│ │ ├── approval.js # Agent 4 — threshold-based approval routing
│ │ └── review.js # Agent 5 — self-correction & pricing verification
│ ├── data/
│ │ └── seed.json # Product catalog, services, FAQs, business info
│ ├── lib/
│ │ └── store.js # In-memory customer visit tracker
│ ├── .env.example # Environment variable template
│ └── server.js # Express server & pipeline orchestrator
│
├── frontend/
│ ├── public/ # Static assets (favicon, apple-touch-icon, OG image)
│ └── src/
│ ├── assets/
│ │ └── bizpilot_icon.svg # Brand icon (header)
│ ├── components/
│ │ ├── ChatPanel.jsx # Message input + conversation thread
│ │ ├── PipelineVisualizer.jsx # Live agent step feed
│ │ ├── DocumentPreview.jsx # Quote/Invoice preview + PDF download
│ │ ├── WhatsAppPreview.jsx # Reply preview + wa.me share button
│ │ ├── StatsDashboard.jsx # Request counts & intent breakdown
│ │ └── InteractionLogs.jsx # Full pipeline history log
│ ├── lib/
│ │ └── pdfGenerator.js # jsPDF A4 document builder
│ ├── services/
│ │ └── api.js # Backend API client
│ └── App.jsx # Root layout, routing, mobile/desktop state
│
├── bizpilot_logo_dark_bg.png # Project logo (README + OG image)
└── start.bat # Windows one-command launcher
sequenceDiagram
participant U as User
participant I as Intake Agent
participant C as Context Agent
participant G as Generation Agent
participant A as Approval Agent
participant R as Review Agent
U->>I: "3 CCTV cameras with installation for Ravi"
I->>C: intent=quote_request, items=[camera×3], customer=Ravi
C->>G: matched products + pricing + loyalty status (returning=false)
G->>A: generated quotation document (total ₹10,350)
A->>R: auto-approved (below ₹12,000 threshold)
R->>R: verify unit prices × quantities, line totals, GST
alt pricing correct
R->>U: ✓ Verified quote + WhatsApp reply ready
else pricing issue found
R->>G: regenerate with correction feedback
G->>A: new document
A->>R: re-verify
R->>U: ✓ Corrected quote delivered
end
BizPilot AI was built for the TakeOver Hackathon — a 24-hour product sprint focused on real-world impact for Indian SMEs.
The problem is real. The people it's built for are real. Every design decision — the 5-agent pipeline structure, the rule-based fallback, the WhatsApp-first output, the multilingual replies, the ₹12,000 approval threshold — came from conversations with actual business owners, not assumptions.
Made with ☕ and urgency in Bengaluru.