We don't talk about AI — we show it running

Seven working solutions

• Hardware-comparison chat. One question, five answers: edge CPU, edge NPU, AX650N, GPU server, cloud — with honest per-run metrics.
• RAG with a knowledge graph. Answers over a Russian legal corpus with citations down to the page.
• Document checklist. A contract or an invoice photo — extracted fields and what's missing.
• Text-to-SQL in Russian. The model writes the query, shows it, runs it — and repairs it after database errors.
• Privacy demo. The same text via three routes: local, cloud as-is, cloud after de-identification — byte-counted.
• TCO calculator. Cloud APIs vs your own server vs your own board, in rubles, with sources for every price.
• AI site editor. Say a remark in plain language — the site fixes itself in a sandboxed preview. Ships on the client's own edge box.

The demos are Russian-language by design: the corpus is Russian law, the contracts are Russian contracts. The lab itself, the approach and the hardware travel fine — if you read English and want the numbers, write us.