On this page
- Products and use cases
- How the platform fits together
- About the Platform
- The Challenge
- Our Approach: AI Where It Can Be Checked
- What We Built
- The Engineering Decisions Behind It
- Safety and Privacy by Design
- Where It Stands Today
- How we work
- Technology
- The 24-month roadmap
- How one referral travels through the suite
- How much each agent may do on its own
- Keeping eleven agents honest
- What can go wrong, and what stops it
- Why the Roadmap Runs Two Years
- The Operating Model
- Cost and Scale
- Decisions to Make First
- Where We Start
- Lessons you can reuse
- Questions
About the Platform
A healthcare software company sells five products on one platform: clinical documentation, dictation, referral tracking, patient texting and payer-rate benchmarking. They share one login, one bill and one audit trail, and a customer can start with a single product and add others later.
We engineered the five products and the AI inside them, so the platform works both as five separate products and as one suite.
The Challenge
Healthcare staff lose their days to work that is mostly reading, re-typing and chasing: charting after the visit, keying referrals in from emails and scans, calling patients who do not pick up, and trying to read payer rate files that nobody designed for humans.
That is the right territory for AI agents, and also the risky one. The data is protected health information, the decisions affect patients, and a confident wrong answer costs more than a slow right one. So the question is not whether to add agents, but where, in what order and with what limits.
Our Approach: AI Where It Can Be Checked
We did not bolt chatbots onto five products. The rule is simple: AI goes where its output can be checked, and everything else stays in ordinary code. Where agents run today, in the documentation product, they follow the rules below, and the roadmap extends them to every product.
- Deterministic core first: The numbers, routing rules and records stay in ordinary code and queries. Agents read, draft and propose on top of them.
- A guardrail at the door: The documentation agents check that input is clinical before anything is drafted, and refuse the rest.
- Bounded toolsets: Each agent gets a short list of things it may do, such as reference lookups for codes, and nothing else.
- Provenance: The drafting agent documents only what the source contains and leaves gaps visibly empty. The roadmap extends this to quoting the exact source line for every extracted value.
- A person decides: Agents draft, flag and propose. A clinician, coordinator or analyst confirms before anything is signed, filed or sent.
- Audit and evals: The roadmap makes evaluation sets and audit logging standard for every agent, so changes are measured rather than guessed.
What We Built
Five products on one platform, each doing one job well.
- Note Studio: A drafting workspace for clinical notes. Paste a transcript, upload a scan or join a meeting, pick a template, and an input guardrail, a drafting agent, a code-check agent and a template builder produce a note for the clinician to review and sign.
- Voice Desk: A desktop dictation app that records the visit, transcribes it with speaker labels and pastes clean text into whichever application is in focus, with voice commands, snippets and a personal dictionary.
- Handoff: A referral workflow from send to acknowledge, accept or decline, with patient details revealed in stages, provider identity checks, reminders and expiry, and rules-based triage that routes each referral.
- Nudge: Two-way patient texting with consent tracking, approved templates, timed reminders, follow-ups when nobody replies and a task queue for exceptions, run by a workflow engine with an audit log.
- Rate Lens: A payer-rate benchmarking workspace that compares negotiated rates by procedure code with peer and market averages by state and payer, with a branded report.
- One platform underneath: Shared sign-in and licensing, per-seat billing, tenant-scoped data and an audit trail across all five.
The Engineering Decisions Behind It
Ordinary code where facts matter: Rates, routing rules, records and workflows are deterministic. Three of the five products need no model at all, and that is the right answer for them.
A guardrail before the model: The documentation agents check that input is clinical before anything is drafted, so out-of-scope text never reaches the drafting agent.
Reference lookups, not model memory: Diagnosis, medication and lab codes come from standard reference services while the note is drafted.
Draft only what was said: The drafting agent never invents or infers facts, and missing fields stay visibly empty so the clinician reviews before signing.
Staged disclosure of patient data: Referral details are revealed in stages as a referral progresses, so each party sees only what the stage requires.
Approved messages and enforced opt-outs: Texts go out only from approved templates, and opt-outs are enforced on the server rather than trusted to the interface.
One foundation for five products: Shared sign-in, licensing, billing and audit mean a customer can add a product without a second integration.
Safety and Privacy by Design
The data here is protected health information, so privacy is part of the architecture, not a checklist at the end. These are the rules we hold every agent to.
- Minimum necessary data: Send a model only what the task needs, and keep identifiers out of prompts when a task does not require them.
- Redaction in logs and telemetry: Logs and traces never carry patient details.
- Tenant isolation: Each customer’s data is scoped on every query, so one customer can never reach another’s.
- Agreements before data: Provider agreements covering protected health information are in place before any such data reaches a model.
- Controls, not compliance claims: We build the controls. Compliance claims belong to the platform’s owner and its auditors.
Where It Stands Today
The platform runs five products on one foundation. 4 LLM agents work in the documentation product, speech models power dictation, and the referral, texting and benchmarking products run on deterministic engines. That spread is deliberate: it keeps AI where it can be checked, and ordinary code everywhere else.
The next step is a shared agent layer and the 24-month roadmap below: 7 more agents, rolled out in the order that is easiest to verify.
How we work
- 1
Map where the minutes go
Start from the people: clinicians, coordinators, the front desk and analysts, and where their time goes in each product.
- 2
Pick one job per product
Choose the one task where an agent saves the most time with the least risk, and leave the rest alone.
- 3
Build the shared layer first
Guardrails, bounded tools, audit, metering and an evaluation harness, used by every agent that follows.
- 4
Ship one agent at a time
Each behind a flag, with an evaluation set, a human review step and a rollback.
- 5
Measure, then widen
Judge each agent on real use against its evaluation set before giving it more to do.
- 6
Operate and keep improving
Run the evaluation loop continuously, review production samples, and retire or retrain any agent that drifts.
Technology
- React and TypeScript
- Node.js and Express
- Python and FastAPI
- Electron desktop app
- Managed cloud LLM platform
- Speech-to-text models
- Cloud document database
- Cloud data warehouse
- Queues and background workers
- PDF report generation