WACRM — WhatsApp CRM with AI Reply Agent

The challenge

Sales and support teams needed a WhatsApp CRM where an AI agent never answers about the wrong order, handles typo-tolerant knowledge-base search, and never leaks internal instructions or API keys in replies.

Approach

Self-hosted WhatsApp CRM (Next.js + Supabase) with a hybrid knowledge search: exact number lookups, pgvector semantic search, Postgres full-text search, and pg_trgm typo tolerance. A short questionnaire writes agent instructions with fixed privacy rules; a leak guard blocks replies containing API keys or internal logic and hands the chat to a human. Login and dashboard screens were redesigned.

Results

• Order, invoice, and ticket numbers are looked up exactly — the agent never confuses records.
• Four-way knowledge search (exact, semantic, keyword, typo) merges into one answer.
• Quality checks score the agent on test questions with RAGAS- and DeepEval-style metrics ported to TypeScript.
• Built-in prompt-injection tests run on every agent change before customers see it.

Metrics

Technologies

• Next.js
• TypeScript
• Supabase
• PostgreSQL
• pgvector
• React
• Tailwind CSS
• Vitest