Bheex: AI customer service SaaS

Bheex is my company: an AI customer service SaaS. Each customer company configures its agents from an admin panel, uploads its documents and serves its own customers over WhatsApp and voice calls. It runs in production on AWS EC2.

Answering on behalf of a company takes more than plugging a language model into a chat. Answers must come from that company's own documents and be citable, no company may see another company's data, and a person must be able to take over the conversation when needed, even in the middle of a voice call.

I am the founder and the AI engineer. I designed and built the whole system and I run it today: the backend, the admin panel, the agent pipeline, document ingestion, the WhatsApp and voice channels, deployment and the security controls.

1. An intent router and an agent with tool calling on Gemini

   Separating intent detection from the conversation lets each message reach the right flow. Those flows are configured from the admin panel, so changing how an agent behaves does not require touching code.

2. RAG filtered per customer on Qdrant, with cited answers

   Every vector search carries the customer filter, so an answer can only rest on documents from the company that is asking. The citation lets anyone check where each fact came from.

3. Security controls as part of the design

   The agent uses tools with real data, so I limited what it can do: read only SQL queries (SELECT only, allowed tables, forced LIMIT and a maximum run time), SSRF protection, secrets encrypted with AES 256 GCM and isolation between customers. A model mistake should never turn into damage.

Three fronts at once. Isolation: several companies share the platform and none may see another's data, not even through the model. Live voice: handing a call from the AI to a person without dropping it means coordinating speech to text, text to speech and the handoff in real time. And tools with real data: SQL queries and outbound requests are attack surfaces that had to be closed up front.

Bheex runs in production today as a multitenant platform on AWS EC2, with automated deployment through Docker CI/CD. The WhatsApp channel (Twilio webhooks) and the voice AI channel are in production: voice uses speech to text and ElevenLabs text to speech, and a person can take over a call without it dropping. Answers cite each customer's documents and the security controls limit what the agent can read and run.

• NestJS 11
• Prisma
• Next.js 15
• BullMQ
• PostgreSQL 16
• Redis
• Docker
• AWS EC2
• Gemini
• Qdrant
• Twilio
• ElevenLabs

Visit the Bheex site