Problem
Teams were drowning in WhatsApp threads, email, and Telegram — inbound messages lost in noise, outbound responses slow and inconsistent. They needed a single AI-powered platform that could handle intelligent messaging across channels, grounded in their own data, with real operational controls.
What I built
- Leading the project end-to-end — database design (PostgreSQL on Neon via Prisma), PR governance, and Cloudflare deployments across the team.
- Self-hosted Coolify infrastructure running microservices: WhatsApp integration, Telegram bot, and email responder, all plugging into the Wondercall AI core.
- Frontend on Cloudflare Pages; backend APIs on Cloudflare Workers with the Hono framework — edge-first performance, minimal cold starts, global reach.
- RAG pipeline with OCR ingestion, hybrid dense + keyword retrieval, reranking, and grounded refusals. Citation-linked answers so users can verify every response.
- Multi-tenant isolation for data, retrieval index, and model-call budgets.
- Eval fixtures and retrieval-quality regression tracking, so degradation surfaces before users notice.
Outcome
Platform in production serving real inbound and outbound messaging workflows. Retrieval quality holds up on messy scanned-PDF corpora, and every AI answer links back to the source span.
Stack
Hono Prisma PostgreSQL · Neon Cloudflare Workers Cloudflare Pages Coolify pgvector RAG
Building something similar?
If you're working on a project that rhymes with Wondercall AI, let's talk. A 30-minute call is free.