Read prescriptions and supplier invoices as structured data, answer questions from your own medical documents, speak with staff and customers out loud, and prove your counter team is safe — under guardrails a model cannot talk its way past, with a pharmacist approving whatever matters, and no patient data ever leaving your building.
Describe the fields you need once — patient, preparation, dosage, herbs, totals — or upload a real document and click the words you want, and the schema writes itself. Every prescription or supplier invoice then comes back as JSON checked against it.
Customer records and segments, campaigns, demand forecasting, competitor monitoring, compliance checks and consultation notes.
Austria-Codex read straight from the licensed data: 99,937 products, 120,260 packs, 61,256 interactions. Search by name, ingredient, PZN or barcode — or search inside what the catalogue says, with AND, OR, NOT and brackets, so “everything licensed for breast cancer that is not a hormone” is a query rather than an afternoon in a tree. Then read one product six ways — how it is used, what is unsafe about it, what it costs, what could replace it and at what conversion factor. Check up to twenty medicines against each other: what reacts, and what is simply in the basket twice. Or review a whole list against one patient — their kidneys, allergies, intolerances and whether they can swallow a tablet — where four catalogue answers become thirty findings, each labelled as coded data, as our own reasoning, or as a paragraph you still have to read, plus a curated reference kept alongside the catalogue for what it does not carry itself: anticholinergic burden, geriatric potentially-inappropriate-medication lists, CredibleMeds, LiverTox, STOPP/START, FRIDs, FORTA, Lifestyle, and CPIC/DPWG pharmacogenomics, its own tab on the same table — maintained by hand from its own add/edit/delete screen, with a CSV or Excel bulk upload alongside the one-by-one form, not a one-time load, and merged additively with the client's own integrated risk database so nothing hand-curated was overwritten. Two more tabs go deeper still, read-only: genotype/phenotype-specific CPIC/DPWG/FDA dosing recommendations, and AMP Tier 1/2 lab-reportable alleles with Meamedica gene panel descriptions, browsable by gene. Findings triage into Act now, Review, Note and Handled tabs rather than one long scroll, worst first, with a single data-gap bar for whatever patient value is still missing instead of a card per gap, and full provenance one click away rather than open on every card by default. It runs as a workflow rather than a screen: reconcile what they actually take, screen it, triage each finding as relevant, already managed, dismissed or waiting on a lab result, record the intervention, who it went to, when to follow up, and what came of it — with an evidence snapshot of identifiers that can re-derive the finding against any later release — then carry it on to a prescriber consultation letter, the prescriber's own feedback per medicine, a printable ELGA-style medication plan with a drag-and-drop dosing timeline, patient counselling and a follow-up date, each exportable as a PDF. A lab report the patient brings in scans straight into it, too — the same OCR and schema-guided extraction already used for prescriptions and invoices, matched against 149 lab analytes across eleven diagnostic groups, with potassium, sodium, creatinine and INR read straight into the patient context the screening itself uses. A patient's smart-scale body-composition reading — screenshotted off their own phone, scrolling and all — scans the same way, into a one-page take-home report styled exactly like the client's own template rather than this app's usual documents; that screenshot can also arrive by email instead of a browser upload, for a deployment reachable by nobody but itself, with the finished PDF emailed straight back — or, for real patient data rather than testing, by Threema Gateway instead, end-to-end encrypted the whole way through a small dedicated service built on Threema's own SDK. The catalogue's German can be read back in English by a model running on this machine, beside the original rather than instead of it, with every number checked across. Licensed content stays on the screen — never saved, exported or used as training material; the review itself, once something is documented, is saved and resumes if you leave and come back.
Upload a recording and it writes out what was said, then reports how much of the call you spoke, what you were asked, and which objections came back unanswered. Compare the calls that went well against the ones that did not — and practise the next one against someone who does not simply agree.
Fourteen levels of engagement read from a health-care professional's own emails and messages, with the words that decided it and the one thing worth doing next. Someone who asked not to be contacted is a stop, not a low score. Accuracy is measured against examples you label yourself, never quoted from a brochure.
Entities and the relations between them, drawn from what an assistant has remembered — every connection keeping the words it came from. See what holds the picture together, what it treats as one subject, where it disagrees with itself, and when each thing happened.
The question, what was chosen, what was rejected and why, who is accountable, and what it led to. Ask what you decided last time a situation like this came up — including the times it went badly.
Staff hold a spoken consultation with a virtual patient who answers only what they are asked — and half the time is hiding a symptom that needs referral. Marked on the questions they asked, whether they caught it, and what they decided. Cases follow your own procedures.
Talk to an agent or a whole team and hear the answer while it is still being written — the first sentence is spoken about a third of a second in. English and German, with the voice model running beside everything else on your hardware.
Turn a Beipackzettel into language a patient can actually act on, at the reading level you choose, with the original always kept alongside.
Encrypted, tenant-scoped retrieval over your own medical documents. Answers cite the pages they came from and abstain when the source does not support them.
Wholesaler and ERP APIs, document stores, internal services — authorize with OAuth 2.0 or a key and each operation becomes a tool your agents can use.
Teams of agents run as a graph that pauses at the steps you gate — a batch release, a patient message — until a pharmacist approves it.
Agents execute in parallel supersteps over a directed graph, with checkpointed state and interrupts that pause for a pharmacist and resume exactly where they stopped. If you know LangGraph, you know the model.
Register any MCP server and bind it to an agent, or let this workspace be the server: your connected systems and extraction schemas are published as MCP tools other clients can call too.
Agents keep what they learn across runs, scored on retrieval rather than dumped wholesale, with an append-only ledger you can replay or roll back.
Run an agent or a team every minute, hour, day, month or year. Days and longer keep their wall-clock time across clock changes, monthly on the 31st lands on the 28th in February, and a run that overruns its next slot is skipped rather than queued forever.
An agent-to-agent registry so agents can delegate to peers — inside this workspace or across a boundary — under the same approval policy as any other tool.
A prescription photographed at the counter, a faxed order, a PDF from a wholesaler. Pages with no text layer are transcribed by a local OCR model first, and you can see exactly what was read before deciding whether to trust the extracted fields.
Upload a document and check the extracted fields against the source, side by side.
An agent reads whatever is attached to the conversation, alone or as part of a team.
Stable REST routes and scoped keys, so an existing system can do the same thing.
Your own vLLM, OCR and embedding services. Prescriptions, patient records and leaflets are processed on hardware you control — never sent to a third-party model.
Retrieval over medical documents requires an MFA-verified session. API keys are deliberately refused there, however privileged their creator.
Rules an agent cannot argue with: decision tables, authorisation matrices and invariants evaluated before a tool runs, not a prompt asking a model to behave. Policies are versioned, signed, tested before publication and provable with a solver — and every decision leaves evidence.
Role-based access, a hash-chained audit trail, and a full trace of every document read and every tool an agent called, kept with the run.
Run it on your own server, keep the medical records where they already are, and give your team agents that can actually read them.
Start free