Easy AI

Easy AI is where these essays came from. A business points it at their own website; it learns the catalogue, answers customers on chat, and connects voice into the same operating layer through AI Call Center. Those interactions can become tickets, leads, and clean customer data. Most of what I know about shipping AI, I learned being wrong here first.

Reliability was the whole job

The hard part was never a clever answer. It was answering the same way on a customer's worst day, ten thousand times over. A customer fires three messages in a row, and a naive bot replies to each fragment; ours holds them for a few seconds and folds them into one, so the model sees a whole thought instead of a stutter. Long conversations get compacted into a short account of the story, the key points, and what the customer actually wants, so the bot stays coherent past the point where a raw transcript would bankrupt it. The tool loop is bounded, and when it reaches the edge of what it can do, it hands off to a human instead of guessing.

The harness is ours

We never hard-wired a model. The general intelligence we rent: every request routes through a fallback chain across several providers, picks a live endpoint ranked by uptime, and retries on the empty answers nobody likes to admit happen, and we have run dozens of models through evaluation to let only a few ever reach a customer. But the frontier does not read Vietnamese the way our customers write it, so the part it gets wrong we run ourselves: our own Vietnamese embedding and reranking models, hosted in-house, are what make retrieval land in the language. We buy the intelligence we can and build the model we can't, and the harness that holds both, the loop and the memory and the gates, is the part we never hand off.

One layer, every channel

Customers talk where they already are: the website widget, livechat, Facebook Messenger across many pages at once, Zalo, WhatsApp, Line, Discord, email, and the phone. AI Call Center connects Caresoft and text-to-speech through the same operating layer; other voice capabilities are scoped per deployment. Orders and catalogue sync in from Shopee, TikTok Shop, and the large local retailers, so the chat system answers about real stock and real prices, not a brochure.

Built for real volume

Tens of billions of tokens a month. In 2025, the first partner in Vietnam to pass ten billion tokens with OpenAI. None of that came from a bigger model. It came from the unglamorous half nobody downloads.