Why we built this
In 2022 we were implementing WhatsApp automation for clients using Wati, Interakt, AiSensy and Gupshup. The consistent pattern: our clients loved the capability, hated the pricing. ₹35,000–₹85,000/month for a seasonal jewellery shop or a 3-chair clinic that would use the platform heavily 3 months a year and barely at all the rest. The subscription model was built for SaaS economics, but SMB cash flow isn't SaaS.
We also noticed that every reseller (and most "platforms") were adding a 20–40% margin on top of Meta's raw messaging price — which is already published and public. We weren't adding anything on top. Customers were paying twice: once for subscription, once for inflated per-message rates.
The hypothesis: build a direct Meta BSP connection, charge pay-per-message at near-cost, make the money on implementation and services.
The architecture
Deliberately boring, deliberately scalable:
- Edge: Nginx + Cloudflare in front of all inbound webhooks. Meta's delivery guarantees are ~99.9% but spikes happen; the edge absorbs them.
- App servers: Node.js (Express + TypeScript), stateless, horizontally scaled behind a load balancer. One deployment per logical region (we started single-region in Mumbai; now 2 regions).
- Queue: Redis Streams for inbound webhook processing, BullMQ for outbound send queues. Each tenant has its own send queue with per-second rate limits that mirror Meta's rules.
- Database: PostgreSQL 16 with tenant-ID partitioning on the big tables (messages, conversations, contacts). Row-level security policies enforce tenant isolation at the DB layer, not just in app code — belt + suspenders.
- Media: S3-compatible storage (Backblaze B2 for cost) for images, docs, voice notes. Signed-URL delivery, 24-hour expiry matching Meta's media retention.
- AI layer: intent classification + response generation run through a small fine-tuned model for languages (Hindi, Hinglish, 8 regional) + a larger general model as fallback. Everything cached aggressively in Redis — same question = same answer, no re-tokenization.
- Dashboard: Next.js (React Server Components), hosted alongside the API. Shared Postgres, read replicas for analytics queries so writes aren't blocked.
- Observability: Grafana + Loki + Prometheus, single alert channel (Slack), on-call rotation inside our 4-person engineering team.
The hard parts
1. Becoming a Meta BSP (not easy, not fast)
Meta's Business Solution Provider program requires technical validation across webhook reliability, template management, number provisioning, and compliance. The checklist is public; the real work is their review cycle — expect 6–10 weeks from first submission to approved status. Being an approved BSP means we price at Meta's wholesale conversation rates, not the retail rates resellers quote. This is the structural reason Akashvanni can be 60% cheaper than Wati on equivalent volume — not a loss leader, actual cost basis.
2. Template approval purgatory
Every WhatsApp marketing message uses a Meta-approved template. Templates get rejected for reasons the public policy doesn't fully document ("promotional content" vs "utility" is judged inconsistently). We built a template-approval pattern library: 40+ pre-validated templates per common vertical (jewellery, clinic, restaurant, real-estate, D2C) that we clone + personalize for each new client. Average approval time dropped from 48 hours to 4 hours.
3. Account-quality management
A WhatsApp business number has a quality rating (High / Medium / Low). Low-quality numbers get throttled; sustained-low gets banned. This is the single biggest risk in WhatsApp automation and the thing most vendors don't talk about. We built automatic frequency caps per contact (max 1 marketing message per contact per 24h), opt-out enforcement, and a quality monitor that pauses sending the moment Meta's quality score dips from High to Medium. No client has been banned on our platform.
4. Multi-tenant isolation
Clinic A's patient data must never leak to Jewellery Chain B's marketing inbox. We enforce isolation at four layers: app-level tenant check on every query, Postgres row-level security policies as a safety net, separate S3 prefixes per tenant for media, and tenant-ID in every log line so a mis-routed webhook is traceable in seconds not days.
5. Compliance from day one
Built with DPDP Act 2023 in mind: India-region data residency (all databases + storage in Mumbai), consent capture on first opt-in with timestamp + source URL, 30-day export + deletion APIs, audit log of every data access, encrypted at rest (AES-256) and in transit (TLS 1.3), and a published Data Processing Agreement any customer can sign.
What's inside the dashboard
- Team inbox — unassigned / assigned / resolved. Multi-agent, permissions-aware, keyboard-first.
- Broadcast campaigns — segment by tag, schedule, A/B test, auto-opt-out enforcement.
- Chatbot builder — visual flow builder for rule-based bots; natural language mode for AI bots (just describe the business, we train on the client's catalog).
- Click-to-WhatsApp (CTWA) ads integration — Meta Ads Manager connects natively, every lead lands in the inbox with ad_id + utm_source attached.
- Analytics — not vanity. Response times, resolution rates, agent performance, campaign conversion, revenue attribution.
- Integrations — Google Sheets (every chat → row), Razorpay (payment links in message), Shiprocket (shipping status webhooks), Zoho CRM, Tally (invoice sync), generic webhook/API.
Pricing we chose, and why
₹2 per outbound message average (varies by Meta's conversation category), ₹0 subscription, ₹0 per seat, ₹999 one-time setup covering Meta Business Manager verification + first 10 templates + onboarding call. That's it. We don't sell "starter / pro / enterprise" tiers. The price doesn't change with volume — it's the same for a single-chair clinic and a 50-store retail chain.
We make money the honest way: on implementation projects (AI chatbot building, integration work, custom dashboards) which we charge for separately.
What it took to get here
| Milestone | Timeline |
|---|---|
| First prototype (internal use at Appzur) | Q2 2022 |
| Meta BSP application submitted | Q4 2022 |
| Meta BSP approved | Q1 2023 |
| First 100 paying businesses | Q3 2023 |
| Multi-region deployment | Q2 2024 |
| AI chatbot layer launched | Q4 2024 |
| DPDP Act 2023 compliance audit passed | Q2 2025 |
| 1,500th paying business onboarded | Q3 2026 |
| 1,000,000th message processed | Q3 2026 |
"We never set out to compete with Wati or Interakt. We set out to make it make sense for our own clients. It turned out their clients had the same problem, which is why 1,500 of them are now ours."
What we'd do differently
We'd build the multi-tenant isolation at the database layer from day one, not retrofit it 6 months in. We'd invest in the template-approval pattern library earlier — it unblocks client onboarding more than any other single thing. We'd hire a dedicated account-quality analyst sooner; the engineering team doing both ops and dev slowed us down in year 2. And we'd document the Meta BSP application process publicly — we've been asked 100 times, and sharing it would help the whole WhatsApp ecosystem.
Want Akashvanni for your business?
Or want us to build a custom WhatsApp / voice bot / AI platform for yours?
Talk to the team →