AI put inside the process that already runs, not on top of it.

AI CRM & WhatsApp Assistant Development

Renting an AI assistant SaaS gets you to demo quickly. It also caps you at what that SaaS decides to build, prices you per contact, and hallucinates prices on the products it never saw. I build the AI CRM that lives inside your business — grounded on your real catalog, with hard guards against making up numbers, on infrastructure you own.

Why it matters

Every hallucinated price is a lost sale, a refund, or a lawsuit. Every locked contact list is a moat someone else built around your customers. Every $2,000/month AI SaaS bill scales with growth in the wrong direction. Owning the pipe fixes all three.

What you get

  • →Anthropic Claude Sonnet 5 (for reasoning) + Haiku 4.5 (for routing) with a complexity-classifier that picks the model per turn — 60–80% cost reduction vs Sonnet-only.
  • →RAG grounded on the live product catalog, so the assistant only quotes prices it actually verified against the source.
  • →Hard guards: the assistant is not allowed to invent a SKU, a price, a stock level, or a promise the business cannot keep.
  • →Unified inbox across WhatsApp Business Cloud API, MercadoLibre v1, TiendaNube v1, Facebook Messenger, Instagram, plus a web widget.
  • →5-role RBAC (owner, admin, agent, viewer, bot) with full audit log per conversation.
  • →Handoff to human — the assistant knows when to escalate, and does so before frustrating the customer.
  • →A 263-test Playwright E2E sweep so a schema change never silently breaks a production flow.
  • →Zero-downtime deploy with schema-drift guard — the AI stays online while the underlying database migrates.

Stack

  • Anthropic Claude Sonnet 5
  • Claude Haiku 4.5
  • OpenAI GPT-5
  • LangChain
  • Whisper
  • Next.js 15
  • NestJS
  • PostgreSQL
  • pgvector
  • Socket.io
  • WhatsApp Cloud API
  • MercadoLibre API
  • TiendaNube API
  • Playwright

Delivered projects using this stack

Frequently asked questions

What model do you use — GPT, Claude, Gemini, Llama?+

Claude Sonnet 5 for reasoning and Haiku 4.5 for routing / simple turns, chosen per turn by a complexity classifier. This routing routinely cuts LLM cost by 60–80% without any measurable quality loss. GPT-5 and Gemini are available as fallbacks when a client requires vendor diversity or when a specific benchmark favors them.

How do you prevent the AI from making up product prices or stock levels?+

The assistant does not have direct access to a price list in its context. Every price question triggers a RAG query against the live catalog (product API or database), and the assistant is instructed at the prompt level to refuse to answer if the catalog does not confirm the SKU. If it cannot verify, it escalates to a human. This is the single most important design decision in a sales AI.

What about WhatsApp — Baileys, Evolution, Chatwoot, Whapi?+

Only WhatsApp Business Cloud API official (via a Business Solution Provider like Gupshup, 360dialog, Wati or Twilio). Baileys, Evolution and every other unofficial method violates WhatsApp's ToS, is banned on any real business number within days, and is a professional non-starter. Meta charges per conversation directly — cheaper than any wrapper.

Can the AI handle voice notes?+

Yes. Whisper transcribes the audio, the transcript enters the same LLM pipeline, and the response goes back as text (or as a synthesized voice note when the client prefers). Predia Digital runs this flow in production for a real-estate booking use case.

Ready to start?

Send a written brief with the business problem to solve. I come back with a concrete plan against that document, in the client's language.

Contact →

Other services