I'm the founder, CEO and chief engineer of Kalorad. We train our own language model from zero and build the assistant that runs on it inside a customer's own network. No calls to anyone else's AI, weights the customer can own outright, and a tokenizer made for Ukrainian.
It's for the people who can't send their documents to a foreign vendor: banks, hospitals, ministries, law firms.
The model. Our first full pretraining run finished in September: 683M parameters, 13.1B tokens, one GPU, $86 of compute. The production model is a mixture of experts with multi-token prediction and 1.3B active parameters. It has a fast mode and one that thinks longer. It's training now on MareNostrum 5 at the Barcelona Supercomputing Center, on 10,000 H100 hours from the EuroHPC AI Factory call. I wrote the training stack (Slurm jobs, DDP and FSDP, the Muon optimizer, torch.compile) and found the bug that had been making training 39% slower.
The product. Chat and an agent that works with Google Workspace, Microsoft 365, KeyCRM, Bitrix24 and email. MCP in both directions. Add-ins for Word, Excel, PowerPoint and Outlook, an add-on for Google Workspace, desktop apps for Windows and macOS, and an OpenAI-compatible API, so anything already built for that API plugs in unchanged.
The team. Three engineers: me, plus two who joined in October 2026. The code is closed. Results go to LinkedIn as they come, including the runs that go wrong.
Most client work is under NDA. These are the public pieces.
Five years of full-stack work for private clients in C#/.NET, Python and TypeScript, two of them leading a team. In 2025 and 2026 I worked on the other side of model training: AI Trainer at Meta, then AI Training Contractor at Scale AI, reviewing and rating model output, mostly code. BSc (Hons) Artificial Intelligence, De Montfort University, First Class.