OutreachX

Scale Your Reach. Not Your Team.

Comment

OutreachX

Scale Your Reach. Not Your Team.

AI-Powered Multi-Channel Campaign Automation — WhatsApp · Voice Notes · AI Phone Calls

Inspiration

Every small business owner in India has faced the same wall: you have a product worth selling, a list of potential customers, and absolutely no time or budget to reach them at scale.

We watched friends running D2C brands spend hours manually sending WhatsApp messages one by one. We saw real-estate agents copy-pasting the same follow-up text to 200 leads every morning. We saw SaaS founders paying ₹15,000/month for tools that only did one channel — and still required a developer to set up.

The breaking point was realising that broadcasting exists, but intelligent outreach doesn't — at least not for the 6+ crore Indian MSMEs who can't afford enterprise tooling.

India has 900M+ WhatsApp users, a booming startup ecosystem, and customers who expect instant, personalised communication. The infrastructure to serve that expectation — multi-channel, AI-driven, autonomous — simply didn't exist at an accessible price point.

That gap became OutreachX.

What It Does

OutreachX is an all-in-one AI campaign automation platform that lets any business launch, manage, and analyse outreach campaigns across WhatsApp, Voice Notes, and AI Phone Calls — from a single interface, in minutes.

Campaign Builder A 7-step wizard guides users from campaign title to live launch: pick channels, upload assets, write a description (AI enhances it), upload a PDF knowledge base, upload contacts via CSV/Excel, preview with live cost estimation, and fire.

AI Content Generation Gemini 2.5 Flash refines campaign copy in your chosen tone. Text-to-Speech converts descriptions into WhatsApp-ready voice notes (OGG/Opus). AI image generation creates campaign visuals on demand.

WhatsApp Automation Bulk campaign sending with text, images, PDFs, and voice notes. Inbound replies are handled by an AI agent that uses campaign context, chat history, and your uploaded PDF knowledge base to respond accurately — 24/7, zero human intervention.

AI Phone Calls via VAPI Outbound AI calls to every contact. The agent holds natural multi-turn conversations, answers questions using your knowledge base, and logs call status (answered/missed) with full transcription and recording. If a customer on a live call asks for a brochure or pricing sheet, the system dispatches it to their WhatsApp in real time — cross-channel, automatically.

RAG (Retrieval-Augmented Generation) Upload product PDFs, guides, or FAQs. OutreachX chunks them, generates 768-dimensional embeddings via Gemini, stores them in Pinecone, and retrieves relevant context in under 100ms whenever a customer asks a question — on WhatsApp or a live phone call. When the RAG system cannot confidently answer a query, it flags it instead of hallucinating. Campaign owners fill the gap via the Organizer Answers panel, which injects the new answer back into Pinecone instantly — closing knowledge gaps dynamically over time.

Prepaid Wallet & Pay-As-You-Go Billing OutreachX uses a Razorpay-integrated virtual prepaid wallet so businesses only pay for what they send — no subscriptions, no surprise bills.

  • Wallet Top-up — fund your campaign wallet anytime via a native Razorpay modal
  • LaunchGuard — blocks campaign launch if wallet balance is insufficient for the selected channels and contact count
  • PaymentGuard — performs real-time deductions during AI call execution and manual follow-ups
  • Pricing: WhatsApp ₹6 · Voice Note ₹9 · AI Phone Call ₹15 per contact

Multilingual Voice Synthesis Sarvam AI integration supports 10 Indian languages — Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, and English. Automatic Unicode script detection routes regional language replies to the correct voice profile. Romanized Indic text (Hinglish, Benglish, etc.) is detected via pre-trained heuristics and handled with matching regional voice output.

Smart Inbox Every WhatsApp reply, across every campaign, in one place. Per-contact thread view with full AI auto-reply capability and multilingual support.

Analytics Dashboard Real-time metrics: call answer rate, WhatsApp engagement score, messages sent, users interacted, and unanswered RAG queries — visualised with donut and bar charts via Recharts.

Onboarding Profile Business context (type, audience, brand style, language, compliance notes) is captured once and used by the AI across every campaign — so responses always sound like you.

How We Built It

Frontend: Next.js 16 (App Router) + React 19 + TypeScript + Tailwind CSS 4 + Framer Motion. SWR for data fetching, Recharts for analytics, Clerk for authentication.

Backend: Node.js + Express server handling WhatsApp send/receive, webhook processing, and AI reply generation — containerised with Docker and deployed on Google Cloud Platform (Cloud Run) for auto-scaling, zero cold-start latency, and managed HTTPS.

AI Stack:

  • Google Gemini 2.5 Flash — primary LLM for text generation, description enhancement, TTS, and WhatsApp AI replies
  • Gemini Embedding 2 — 768-dimensional document embeddings for the RAG pipeline
  • OpenAI GPT-4o-mini — agent orchestration inside LangGraph
  • LangChain + LangGraph — agentic RAG pipeline with a Retrieve → Respond state machine
  • VAPI.ai — outbound AI phone call orchestration
  • Sarvam AI (bulbul:v3) — Indian language TTS for regional voice notes and STT

RAG Infrastructure: PDF extraction → text chunking (500 chars, 100 char overlap) → Gemini Embedding 2 (768 dimensions) → Pinecone vector DB (namespace per campaign) → cosine similarity search (top 3 chunks, <100ms)

Billing Infrastructure: Razorpay Payment Gateway → virtual wallet balance stored in Firestore → LaunchGuard checks balance before dispatch → PaymentGuard deducts per contact in real time → full transaction log persisted per campaign

Infrastructure & Deployment:

  • Google Cloud Platform (Cloud Run) — containerised backend deployment, auto-scaling, managed HTTPS
  • Google Cloud Artifact Registry — Docker image storage and versioning
  • Firebase / Firestore — real-time DB, analytics persistence, inbox and wallet state
  • Cloudinary — asset and audio CDN
  • Twilio — SIP voice infrastructure
  • WhatsApp Business Cloud API — messaging channel
  • LiveKit — real-time voice streaming
  • FFmpeg — WAV/MP3 → OGG/Opus conversion for WhatsApp

Challenges We Ran Into

WhatsApp Business API sandbox limits The free tier only allows messaging pre-registered numbers. We architected the entire campaign flow to work within a single registered number during development, while keeping the codebase fully scalable for production multi-number deployments.

Audio format conversion for WhatsApp WhatsApp voice notes strictly require OGG/Opus encoding. Gemini TTS returns WAV/MP3. Getting FFmpeg to run reliably inside a Cloud Run container, produce valid OGG files, upload to Cloudinary, and deliver on time before campaign send — in the right sequence — took significant debugging.

RAG latency under concurrent load Pinecone retrieval is fast in isolation (<100ms), but when combined with embedding generation, LangGraph state transitions, chat history loading, and Gemini inference — all happening per inbound WhatsApp message — the pipeline needed careful async optimisation to stay responsive.

Razorpay webhook reliability Ensuring wallet top-ups reflected accurately in Firestore — even under network failures or duplicate webhook deliveries — required idempotent transaction handling and signature verification on every Razorpay event.

VAPI + LangGraph agent context injection Injecting dynamic RAG context into a VAPI call's system prompt before the call connects required pre-fetching document chunks at call initiation time and constructing a fully-formed prompt — a flow that had no clean documented path and required custom engineering.

GCP Cloud Run cold starts The Express backend occasionally hit cold start delays during low-traffic periods. We resolved this with minimum instance configuration and container pre-warming to keep the WhatsApp webhook response window under Meta's 20-second timeout.

Multi-turn conversation coherence Maintaining coherent WhatsApp conversations across sessions meant loading the last 20 messages, the campaign description, and RAG context simultaneously — while staying within Gemini's context window and keeping response time under 3 seconds.

Accomplishments That We're Proud Of

  • End-to-end RAG pipeline working in production — from PDF upload to live AI responses on WhatsApp and phone calls, with sub-100ms vector retrieval
  • Self-healing knowledge base — the Organizer Answers feedback loop means the AI gets smarter with every unanswered query, without any retraining
  • AI phone calls that actually feel human — multi-turn conversations with Sarvam + VAPI that surprised us during testing
  • Full campaign lifecycle in one platform — from a blank canvas to a live multi-channel campaign (WhatsApp + voice note + AI calls) in under 5 minutes
  • Zero manual follow-up — the Smart Inbox auto-replies to every inbound message using the full campaign + knowledge base context, completely autonomously
  • Real billing, not a demo — a live Razorpay prepaid wallet with LaunchGuard and PaymentGuard working end-to-end in the prototype
  • Production deployment on GCP — containerised backend on Cloud Run with auto-scaling, not just a localhost demo
  • Multilingual outreach — Sarvam AI enables voice notes and calls in 10 Indian languages, making OutreachX viable for Tier-2/3 city businesses

What We Learned

  • RAG is only as good as your chunking strategy. The 500-character chunk size with 100-character overlap wasn't arbitrary — smaller chunks lose context, larger chunks dilute relevance. Getting this right made a measurable difference in response quality.
  • WhatsApp is infrastructure, not a channel. For Indian businesses, WhatsApp is the CRM. Building natively on top of it — not just adding it as a feature — is the only way to build something people actually use.
  • Billing is a product feature, not an afterthought. Integrating Razorpay taught us that payment flows need the same level of UX care as any core feature — a clunky top-up flow kills conversion as fast as a buggy campaign builder.
  • LangGraph forces you to think in states, not functions. Switching from sequential LangChain calls to a proper state machine changed how we reasoned about the AI pipeline — and made debugging dramatically easier.
  • Cloud Run changes how you think about deployment. Containerising the backend and deploying to GCP forced better discipline around environment variables, port binding, and graceful shutdowns — habits that will outlast this project.
  • Voice UX is unforgiving. A 2-second pause in a phone call feels like an eternity. Optimising VAPI response latency taught us more about real-time systems than months of backend work.
  • Agentic systems need guardrails from day one. An AI that can send WhatsApp messages and make phone calls to real people needs careful prompt design and fallback handling — we learned this the hard way during early testing.

What's Next for OutreachX

Short-term

  • [ ] Production WhatsApp Business API (verified business account, unrestricted messaging)
  • [ ] Campaign scheduling — queue campaigns for optimal send times
  • [ ] A/B testing for campaign descriptions and tones
  • [ ] Contact segmentation and tagging

Medium-term

  • [ ] CRM integrations — HubSpot, Zoho, Salesforce
  • [ ] Instagram DM and Telegram channel support
  • [ ] Deeper analytics — conversion tracking, reply sentiment analysis, funnel visualisation
  • [ ] Team collaboration — multiple agents managing the same inbox
  • [ ] Subscription tiers on top of the prepaid wallet for high-volume senders

Long-term vision

  • [ ] OutreachX as an API — let any SaaS embed AI outreach into their product
  • [ ] Become the default AI outreach layer for Indian businesses — the "Salesforce for Bharat"
  • [ ] Vertical-specific templates: real estate, D2C, edtech, healthcare, financial services
  • [ ] Full GCP-native deployment with Cloud Spanner, Pub/Sub for event-driven campaign triggers, and Vertex AI for model fine-tuning

The goal isn't to be another marketing tool. It's to give every Indian business — from a 2-person startup in Patna to a 200-person team in Pune — the same outreach intelligence that enterprise companies pay crores for.

Built With

  • antigravity
  • firebase
  • gemini-api
  • gemini-embeddings
  • gemini-tts-api
  • google-cloud
  • langchain
  • langgraph
  • nextjs
  • pinecone
  • razorpay
  • twilio
  • vapi-ai-agent
Share this project:

Updates