Nothing you type, record, or upload here is sent to Tathastha Labs or any server.
A named-entity model finds people, organizations, and locations in raw clinical text — patient and clinician names, facility names — the same category of model that underlies de-identification and structured-extraction pipelines.
~45 MB first run only
This is a general-purpose vision classifier, not a diagnostic model — we're showing the mechanics of an in-browser imaging pipeline (upload → preprocess → inference → ranked output), the same shape as the modality-specific models we build for imaging engagements.
~90 MB first run only
An open-source biomedical NER model, fine-tuned specifically to recognize disease and condition mentions in clinical text — the kind of model that feeds structured problem lists and registries.
~65 MB first run only
Speech is transcribed first, then a small instruct model drafts a structured clinical summary from the transcript — the same two-stage pattern behind our ambient scribe work, running entirely on-device here. Both models are small for browser-friendly size, so treat the formatting as a rough draft, not a polished SOAP note — production engagements use larger, evaluated models for that.
~340 MB first run only
Text detection finds and reads every line on the page, in the language you choose, then a small instruct model turns the raw text into structured fields — vendor, date, totals. The same detect-then-structure pattern behind document-processing pipelines, running entirely on-device here.
~20 MB first run per language, plus the drafting model used above
Text is split into passages and embedded, then a question is matched against the most relevant passages and answered from them alone — the same retrieve-then-generate pattern (naive RAG) behind most document Q&A systems, indexed and queried entirely on-device.
Plain text only, 1 MB max. ~25 MB first run for the embedding model, plus the drafting model used above. Larger documents take longer to index.