A few months ago, we were running a local customer support tool under a .nl domain. It did what most modern "AI chatbots" do: it ingested some docs, sat on a page, and answered basic FAQs.
But we had to be honest with ourselves: standard chatbots are a weak abstraction. If a customer texts support, they don't want a paraphrased FAQ link. They want their order cancelled, their refund initiated, or their shipping label generated.
"Chatbot is the door, but the Agent is the room."
When we decided to pivot to the international market with Nousu.chat, we didn't just translate our UI. We completely ripped out our architecture and rebuilt the entire runtime for raw performance on the serverless edge.
Here is how we did it, why we rejected the traditional database tax, and how serverless edge infrastructure allowed us to kill the "per-ticket" pricing model
- The Raw Block Architecture: Ditching the Fixed-Cost Rot
Most AI platforms start by stacking heavy container-based runtimes, expensive vector databases, and massive relational database wrappers. This is fine for venture-backed startup theater, but it creates massive fixed-cost rot and high latency
We built our engine on a "Raw Block" philosophy:
- The Runtime: We migrated the entire agent execution layer to Cloudflare Workers. Cold starts are dead (<5ms execution), and we can run our agent logic globally right next to the user
- State & Caching: Instead of querying a heavy database for every turn, we utilize Upstash Redis and Queues to manage session states and handle high-throughput event buffering
- The Database: When relational persistence is actually required, we route queries to PostgreSQL on Railway. Optimized SQL queries over lightweight pools.
By moving everything to serverless edge infrastructure, our average round-trip API latency dropped to under 50ms (excluding the LLM generation itself)
- Real Resolutions vs. FAQ Spam (The API Hook Runtime)
If your "agent" can't write to your database or execute external APIs securely, it's not an agent; it's a glorified search bar
To make Nousu.chat a true operator, we built an ultra-fast tool-calling router. When a WooCommerce or custom merchant API hook is triggered, the Worker doesn't wait for a heavy backend server to spin up.
Instead:
- The edge worker intercepts the user request
- It evaluates intent and maps it to a strict TypeScript-defined tool schema
- The worker directly executes the secure outbound fetch request to the merchant's API (e.g., to process a refund)
State is updated instantly, and the user receives a confirmation of the action, not just text
Killing the "Ticket Tax"
Traditional enterprise support tools (like Intercom or Gorgias) levy a 'tax' on every single resolved ticket. When you scale, your bill scales exponentially, even though database writes and serverle
ss compute costs are practically static at scale
Because our serverless edge stack runs so incredibly cheap, we don't have to charge a "ticket tax." Operating Nousu at high-scale costs us fractions of a cent per session, which allowed us to structure our business around flat-fee pricing.
What We Learned Going Global
Pivoting from a local market to a global platform required strict internationalization and a zero-tolerance policy for latency. If you're building in the AI space today, our advice is simple:
Stop wrapping slow abstractions. Ditch the bloated libraries, move your runtime to the edge, and focus on building systems that actually do things instead of just talking
Check out how we are running our edge machine at Nousu.chat
Top comments (0)