←

Hader AI

The AI agent that answers, qualifies and closes inside WhatsApp Business, in three seconds, in ninety-two languages, twenty-four hours a day.

hader.ai

Every WhatsApp, answered.

The AI agent that answers, qualifies and closes inside WhatsApp Business, in three seconds, in ninety-two languages, twenty-four hours a day.

live · multi-tenant SaaS · WhatsApp Business · AI agent

  • 3.2sMedian reply
  • 92Languages
  • 24/7Always on
  • 1,450+Commits since April 2025

Product manager and tech lead

I ran Hader from the first spec to a live product: the roadmap, the architecture, the team, and a good share of the code. I shipped features in the API, the portal and the mobile app, reviewed every pull request, and ran the deploys.

  • Product spec and roadmap
  • System architecture and tenancy model
  • API, portal and Flutter app features
  • Code review, releases and server deploys

The team I led

  • WhatsApp specialist

    Meta Business setup, WhatsApp Cloud API, templates and coexistence, green-tick accounts

  • AI engineer

    The bot engine: grounding on tenant data, intent detection, translation, guided order flows

  • Automation engineer

    Broadcasts, drip sequences, catalog imports and syncs, scheduled jobs on BullMQ

  • Full-stack developer

    Next.js portal, Fastify endpoints, Prisma migrations, the operator inbox

Not a chatbot. A command center.

Everything Hader does across WhatsApp, Instagram and Messenger, in one workspace the tenant's team actually likes.

  1. Dashboard section of hader.ai
    01

    Dashboard

    Conversations, median reply, orders and the share handled by AI, with the week's traffic and top products at a glance.

  2. Shared inbox section of hader.ai
    02

    Shared inbox

    Every chat from every channel lands in one thread. Hader reads the catalog, FAQs and policies, answers in the customer's language in about three seconds, and a human takes over with one tap.

  3. Analytics section of hader.ai
    03

    Analytics

    Reply rate, resolution time, busiest hours, revenue by conversation. Forty-plus metrics, live, with CSV and API exports.

  4. Customer profiles section of hader.ai
    04

    Customer profiles

    A living profile for every number: tags, order history, lifetime value and an AI-written persona of what each customer likes and when they buy.

  5. Orders section of hader.ai
    05

    Orders

    A guided flow, not free-form guessing, collects each item, confirms the total and creates a real order the moment the customer says yes. Live statuses, POS sync.

  6. Bookings section of hader.ai
    06

    Bookings

    Clinics, salons, restaurants and agencies. Hader checks availability, books the slot, sends reminders and handles reschedules inside the chat.

  7. Broadcasts and templates section of hader.ai
    07

    Broadcasts and templates

    Write a Meta template once, personalise it per customer, preview exactly how it lands, send to thousands from the tenant's own number. Opt-outs honoured automatically.

  8. By industry section of hader.ai
    08

    By industry

    E-commerce, F&B, real estate, clinics, education and B2B. Each gets its own AI trained on a real business, with a live bot to talk to.

  9. Eight roles section of hader.ai
    09

    Eight roles

    The receptionist, the translator, the closer, the scribe, the dispatcher, the scheduler, the escalator and the night shift. The role shapes itself around the business.

  10. Ninety-two languages section of hader.ai
    10

    Ninety-two languages

    Detects, translates and replies mid-thread with no plugin. Every message is stored in both languages, so the team and the customer read the same conversation in their own words. Fourteen right-to-left scripts.

  11. Live demo section of hader.ai
    11

    Live demo

    Four scripted scenarios on the site: reorder in Arabic and English, book a clinic, where is my order, and an invisible human handoff, with latency, intent and sentiment shown live.

  12. Pay per conversation section of hader.ai
    12

    Pay per conversation

    A conversation is a 24-hour window with one customer. Basic, Pro and Enterprise plans by volume, never per seat. SAR and AED billing.

How each piece was built

The platform is one Fastify API, one BullMQ worker and one Next.js portal over a single Postgres, with 47 route modules, 62 portal pages, 79 models and 97 migrations. These are the subsystems that matter and how they work.

  1. 01

    The bot engine

    Every reply starts by gathering the tenant's own data, then hybrid retrieval (trigram sparse plus 1536-dimension dense embeddings, fused with reciprocal rank), then a sectioned prompt split for provider caching, then a plan-routed completion. The output passes a ten-step validator pipeline and a grounding gate before it is allowed out, and every reply writes a provenance row saying which facts it used.

  2. 02

    Scripted flows and fast paths

    Two layers that skip the LLM entirely: a deterministic node engine for the flows a tenant wants exact (with an optional small-model distress pre-check), and regex fast paths for the questions that never need a model.

  3. 03

    Operator inbox

    One thread list for WhatsApp, Messenger and Instagram. History with presigned media, notes, tags and canned replies. An operator reply self-assigns the thread and pauses the bot. Live updates over server-sent events with single-use nonces, plus tenant-defined if-this-then-that rules evaluated on every inbound message.

  4. 04

    Carts, orders and bookings

    The draft cart is built by parsing the bot's own "added N × item" sentences, then promoted to a real order on a cart marker. Bookings resolve a spoken slot against computed availability. Payment links are minted per order by a per-tenant provider switch, and three signature-verified webhooks converge on one idempotent mark-as-paid.

  5. 05

    Broadcasts and sequences

    CSV, segment or manual audiences are fanned out to one row per recipient, then sent under a per-number token bucket with opt-out, send-window and wallet gates. Drip sequences run on a scheduled tick.

  6. 06

    CRM and ingestion

    Contacts with order history, lifetime value and AI-written personas. Catalog comes in by streaming CSV and XLSX import, cron and webhook connectors, a Shopify scrape-review-commit flow and a Playwright crawler, all converging on one org-scoped upsert so the bot sees a catalog change on its next reply.

  7. 07

    Tenancy and row-level security

    One seam: every authenticated request runs inside a transaction that sets the database role and the current organisation, and Postgres row-level security does the rest. There are exactly two escape hatches, both gated to head-office admins, and a tenant-isolation test that must pass before any release.

  8. 08

    Auth, wallet and billing

    Short-lived access JWT with a rotating httpOnly refresh cookie bound to the device, three roles plus a head-office flag, TOTP, and impersonation for support. A prepaid wallet in integer micro-dollars is charged at send time, with a separate monthly AI-message allowance and dunning.

  9. 09

    Voice, read API and eval

    The platform is the brain for an external voicebot: it compiles a budgeted prompt and form config and ingests the call lifecycle. A public read API with per-key rate limits and HMAC-signed outbound webhooks. An eval harness runs golden sets through the real engine with deterministic scorers and an LLM judge.

The WhatsApp and Meta side

The product lives inside WhatsApp Business, so the Meta integration is the heart of it, and the part with the most rules.

  1. Cloud API, multi-number

    Each tenant connects one or more official numbers. Every channel carries its own credentials, encrypted at rest with AES-256-GCM, and its own bot on-off switch.

  2. Webhooks

    Inbound traffic arrives on an HMAC-verified webhook, resolved to a channel by phone-number id. The handler dedups, handles STOP, upserts the contact and thread, stores media, and dispatches to the bot. Delivery statuses flow back into broadcast counters and wallet refunds.

  3. The 24-hour window

    Meta only allows free-form replies within 24 hours of the customer's last message. The platform enforces the window itself and surfaces it honestly to operators rather than letting a send fail at Meta.

  4. Templates

    Broadcasts outside the window use approved Meta templates. The template list is synced from Meta, retired templates are shown as missing, and a template send is previewed exactly as it will land, with the Meta cost shown before sending.

  5. Media

    Incoming audio and images are fetched from Meta, transcoded where needed and cached, with presigned links that expire rather than stored URLs.

  6. Coexistence and Embedded Signup

    Normally connecting a number to the Cloud API means giving up the WhatsApp Business app on the phone. Coexistence lets the business keep its handset and connect the same number. The connect flow trades Meta's 30-second authorisation code for a token, proves the phone number, cross-checks credentials, writes the channel before subscribing the webhook, then requests contacts and, only with explicit consent, up to 180 days of history.

  7. Handset echoes

    When the owner answers a customer from their phone, Meta echoes the message to the platform, and the bot stops talking over them. This is the stream that makes coexistence usable day to day.

  8. Compliance and the green tick

    Meta Business Partner setup, App Review for the embedded-signup permissions, SOC 2 and GDPR posture, and official-business green-tick accounts handled per tenant, with the Meta-side traps written up so the next connect does not repeat them.

How it is built

A multi-tenant SaaS: each tenant loads its catalogue, hours, FAQs and policies, connects its official Meta WhatsApp number, and the assistant answers that business's customers using only that tenant's own data.

API
Node 22, Fastify, TypeScript. All business logic, Meta webhooks and the bot engine.
Worker
BullMQ on Redis 7. Imports, syncs, broadcasts, drip sequences, scheduled ticks.
Portal
Next.js 15 App Router with Tailwind. Tenant portal plus the ALIGNED head-office console.
Data
PostgreSQL 16 behind PgBouncer, Prisma, row-level security per tenant.
Mobile
Flutter for iOS and Android, a pure client of the same API as the web portal.
Channels
WhatsApp Cloud API, Messenger, Instagram, phone lines. Shopify, CSV and REST catalog imports.
Infra
pnpm workspaces and Turborepo, native systemd units behind Caddy, pull-based deploys with health checks and automatic rollback.
Quality
An end-to-end suite with a tenant-isolation gate that must pass before any release.

It speaks like a person who has read the room.

Hader writes like the business’s best front-desk colleague: observant, direct, fast. One customer, one thread.

  • Specific.not general.
  • Warm.not friendly.
  • Direct.not curt.
  • Curious.not eager.
  • Brief.not cryptic.
  • Multilingual.not untranslated.