Agentic AI & data platforms

Useful assistants act: they search a library, run a calculator, fill in a form, then explain. Doing that reliably with affordable models is an engineering and research problem. What we work on: • Tool use with small and free models: searching before the model answers (small models often skip tools), requiring the calculator when a question asks for an amount, and turning tool calls that a model writes as text into real calls. • Model routing: each feature has a primary model and fallbacks; providers that hit their quota are paused, withdrawn models are benched, and answers still stream when one provider fails. • Structured outputs people can use: draft tax declarations laid out like the official forms, and an export in the tax authority's own XML format, validated against its published schema, with personal details filled in on the user's device only. • Data pipelines: collecting new official documents every few hours, OCR with a per-file cache so only new files are processed, and pushing reviewers' decisions to the live search index without a rebuild. • Channels: the same assistants on the web and on Zalo, with replies formatted for each.