Ru
En

Artificial intelligence: in the products and in the lab

We apply language models where a human checks the result, and we research architectures that have not become a product yet. These two lines are deliberately kept apart: working mechanisms live in the products, hypotheses live in the lab.
THIS TEXT WAS GENERATED BY AI
This text was generated by artificial intelligence. The draft of the page was written by a language model and then checked and approved by an editor — by the same rule the page describes: the machine prepares the draft, the human makes the decision.
Our approach
A language model is good at taking away routine and bad at carrying responsibility.
That is why one rule applies across all our products: the machine prepares the draft, the human makes the decision. We do not ship features where a model’s answer goes to the user unchecked — and we do not call something intelligent when it is not.
What already works in the products
Standard tasks in Pipir

Most of the tasks are written by methodologists by hand from the best problem books. The language model is used only for standard tasks, where a volume of uniform variants is needed, and its draft is approved by a methodologist through a separate interface.

pipir.online
Bots under parental control

In Family chat a bot, including an AI assistant, appears in chats only after a parent has approved it. The control is implemented at the server level rather than in the app settings.

gaml.tech
Tooling for operator assistants

The payment platform exposes its functions through the standard protocol for connecting tools to language models. Operator assistants are assembled without a separate integration for every scenario.

4pay.online
The research part: the engAIne project
A typical application built around a language model is simple: a request goes to the model and the answer comes back to the user. Anything that looks like memory or reasoning is imitated by inserting text into the request.
In the lab we are testing a different approach. engAIne is a cognitive architecture in which every function is a separate, continuously running process with its own state: perception, memory with search by meaning, attention management, a working space, and consolidation of accumulated experience into stable knowledge. The system is assembled into a process tree that tolerates the failure of individual parts. Local models and cloud providers work together: simple requests are served quickly and locally, complex ones go to the large models.
Status of the research
engAIne is a research project: we are checking whether such an architecture behaves more reliably than an ordinary wrapper around a model on long tasks.
Some of the layers are fully implemented, some are still evolving; the direction is not offered as a finished product. If the hypothesis holds, the work will become the basis of the intelligent features of our products — from an operator assistant to document processing.
Discuss working together
If you are interested in the engAIne architecture, in applying language models in your product or in joint research — write to us.
The direction is run inside the lab and is not offered as a finished product, but we are ready to discuss a task or a joint test of a hypothesis.

Write to us
Learning tasks → pipir.online
Bots under parental control → gaml.tech
Payment platform → 4pay.online
What has changed over the past months → blog, category «Dva PiEr.Lab» (in Russian)