Our own Ali-E AI: why we're building an in-house LLM

Today Ali-E's AI features run on strong external models — that let us deliver real value to users fast. But strategically we're building our own AI. Here's why.

Three reasons to move off third-party AI

Cost: every call to an external model costs money, and as the user base grows this becomes a significant expense that pushes on the price for the customer.

Privacy: business data is sensitive. An in-house model means more control over where and how data is processed.

Control and specialization: a general model 'knows a bit about everything'. A model trained on the language of sales, inventory and finance can be more accurate and predictable in exactly our scenarios.

How we do it

We assembled a curated set of training examples for the product — dozens of categories across eight languages, including business scenarios and correct refusals. From it we expanded the dataset to thousands of examples and fine-tuned an open-weight model (via LoRA) on accessible hardware.

In parallel we already shipped a knowledge layer (RAG): the model blends your company's relevant data into the answer, so it responds from facts, not 'the average of the internet'.

The transition plan

We do this step by step and without risk to the user: first RAG and data quality, then our own model as a service, then a gradual replacement of external calls. If something goes wrong, the system falls back to proven mechanisms. No degradation of experience.

What it gives you

Over time — lower, more predictable AI cost, better privacy and an assistant that understands business tasks better and better. We're building our own technology, not a wrapper over someone else's API.

Follow the blog: we share progress openly. It's a rare chance to watch a product grow its own intelligence.