Nexunia: AI agents that execute, audited at every step
Over 8 years designing AI architectures led me to Nexunia: a multi-tenant SaaS platform where a language model interprets the conversation but never touches a system directly. Every business action goes through a proprietary tool gateway, exposed as an MCP server.
The pieces that make this trustworthy
Tool Gateway as an MCP server
The model decides which tool to call; the MCP server validates the session's tenant, executes against the real backend, and returns a structured result. The model never sees third-party credentials or executes anything directly.
Multi-tenancy with RLS from day one
A single shared database, with PostgreSQL Row-Level Security enabled from the very first migration — not as a later improvement. The customer identifier is always resolved from the authenticated session, never from the request.
Booking Service with no double-booking
An appointment only exists once the backend returns a real identifier, never before. The flow goes through a temporary hold and an atomic transaction — explicitly tested with simultaneous requests for the same slot.
RAG limited to FAQ, never pricing or availability
Semantic search over documentation answers frequently asked questions. Pricing and availability always live in structured tables and are queried in real time — never generated as free text.
Full audit trail for every tool
Every execution is logged: what was called, with what parameters, what result the backend returned, in which conversation and for which customer. Queryable from an admin panel.
Human handoff at any point
A human can take control of any conversation — over WhatsApp, voice, email or web chat — and hand it back to the AI at will, without losing context of what was already executed.
The same conversation, any channel
Specialized agents, not just a chatbot
Nexunia's roadmap includes a full catalog of specialized agents per module. Here's what already has a real backend running, versus what's still on the roadmap:
With real backend
- Omnichannel chat (WhatsApp, web, voice, email)
- CRM
- Social media (beta, real publishing on X)
- Web management
- SEO
On the roadmap
- Ecommerce (Shopify/WooCommerce)
- Marketing & loyalty
- Google Ads
- Accounting & finance
Guards matter as much as features
In voice especially, a language model can sound confident without being right. That's why the agent runtime includes a growing set of mechanical anti-hallucination guards — hardened one by one after investigating real incidents with production data, not just in theory.
Want to talk about how this gets built?
If you're evaluating building an AI agent that executes real actions, let's talk architecture, not just prompts.
Let's talk