A reference design · For physician-led practices · 2026

An AI operating layer for the modern practice.

A custom AI layer that handles the small clinical and administrative tasks bogging down your day — running on hardware in your building, not in someone else’s cloud.

A working sketch

Offload tasks in seconds.

One Assistant that knows your charts, your inbox, your schedule, and your renewals. Ask in plain English. Get a structured proposal. Approve or reject. Done.

On-device · cedar-rig.local

Ozempic PA · S. Kim

MC
Dr. Chen9:14 AM

Draft the Humana PA for Mrs. Kim’s Ozempic. Last approval expired 4/30, she’s due for renewal.

Practice Assistant9:14 AM

Pulled the chart and built a draft. Step-edit requirements look satisfied. Here’s the PA letter.

Prior Auth · Humana · GLP-1 (Ozempic 0.5mg)DRAFT READY
PatientSusan Kim
DOB04/12/1968 · 57F
DxType 2 DM (E11.9)
BMI34.2 (4/30)
Last A1c7.8% (4/30)
InsuranceHumana Commercial
Last PAApproved · expired 4/30
Chart Context

Tolerated well at 0.5mg.
Hypoglycemia symptoms documented on prior sulfonylurea trial.

Drafted Justification

FormHumana GLP-1 PA
Step therapyMetformin 12mo · documented
FailureSulfonylurea hypoglycemia
Fax to1-877-XXX-XXXX
Formulary criteria satisfiedCited §4.2b · Humana commercial 2026

Letter cites her A1c, BMI, completed step therapy, and the documented sulfonylurea failure. Ready for your signature.

Draft Prior AuthBuild a PA letter from the chart, ready to sign.Open
Refill ReviewPull med list, last labs, and pending requests.Open
Patient MessageDraft a reply in your clinical voice.Open
Referral LetterCompose a referral from the visit note.Open
Voice NoteCapture a quick dictation between patients.Open
Message Practice Assistant…

Where time goes today

Two stacks. No shared view.

The practice runs on two separate piles of work. The business of the practice on one side, the clinical overhead on the other. Each one is solvable. The cost is the four hours a day you spend moving between them.

The business of the practice

Staff schedulinggroup texts
License & DEA renewalsa sticky note
Credentialing paperworkpile of faxes
CME trackingsomewhere
Vendor & supply orderssomeone’s inbox
Reviews & inquirieswhoever has time

The clinical overhead

Prior authsthree insurance portals
Labs to signinside the EHR
Patient messagesthe portal inbox
Refill requestsalways growing
Referral letterson your evening
Voice notesbetween visits, untranscribed

The first stack has nothing to do with patients. The second runs on hardware that never leaves your building.

Why this matters

Your hours should go to patients. AI handles the rest.

Your hours should go to your patients, your partners, and the work only you can do. Not the dozen small tasks that fill the day between visits. AI is what takes those off your plate — without sending a single chart to a cloud you don’t control.

What we build

One layer across the systems you already use.

Two tracks of inputs — the BAA-free business stack and the local-only clinical stack — both feed the rig. The rig produces the small set of daily and weekly artifacts your team will reach for.

01Your existing systemsTwo tracks, one layer

Business · BAA-free

Staff calendarGoogle / 365
Credentialingtracker
Vendor inboxGmail
Reviews & inquiriesPhase 4

Clinical · stays on the rig

EHRAthena / Kareo
Patient portalmessages
Lab inboxresults
PA portalspayer
Voice capture5 surfaces
ClearinghousePhase 3
02The Practice AI layerRuns on the on-site rig

Reads, drafts, classifies, summarizes, routes, escalates — all on hardware in your building.

No cloud BAA requiredModels run locallyApproval on every writeAir-gappable by design
03Outputs & ActionsAI gets things done
Morning clinical briefdaily · 7:30am
Prior auth queuedaily, urgency-flagged
Drafted patient repliesdaily, your voice
Credentialing alertsas needed
Billing denial digestPhase 3
Care gap outreachPhase 4

The layer runs on locally-hosted Llama, Qwen, and Whisper models on the rig. EHR, lab inbox, and patient portal connect through their existing APIs. Patient data does not leave your network. Nothing is written without explicit approval.

The hardware

The Rig. Quiet enough for the back office.

We deploy a dedicated AI workstation to the practice. Always on. Models, context, and memory all run locally. We configure it, harden it, and maintain the software layer through a secure channel that gives us no access to patient data.

Option ARecommended

The Quiet Box

Mac Studio M4 Ultra

  • Memory192GB unified memory
  • ModelsLlama 3.3 70B, Qwen 2.5 72B, Whisper Large
  • Throughput≈25–40 tok/s sustained
  • Power≈60W, fanless under load
  • FormSits under the front desk
  • CostOne-time · mid five figures
Option B

The Workstation

Custom GPU workstation

  • Memory48–96GB VRAM
  • ModelsLlama 3.1 405B (quantized), Mistral Large
  • Throughput≈60–100 tok/s sustained
  • Power≈300W, server-fan acoustics
  • FormServer closet or back office
  • CostOne-time · high five figures

One-time capital cost. No per-token charges. No cloud dependency. No vendor that could change their terms next quarter. The rig is yours.

Voice capture

Capture the encounter however you work.

Audio captured anywhere flows back to the rig over your network, transcribed locally with Whisper, and routed into chart notes, dictation summaries, or the daily roll-up. Five surfaces. One pipeline. Off by default in exam rooms.

Mobile

Phone, between visits.

A one-tap recorder on iOS / Android. Dictate a note, capture a reminder, summarize a hallway decision. Streams encrypted to the rig, transcribed locally.

Tablet

iPad at the bedside.

Open the chart, tap record, dictate the note. The agent structures it into your SOAP template and routes it back to the EHR for sign-off.

Laptop

Desk-bound composition.

Live transcription in a side panel while you write. Reference dictations while drafting referral letters or PA narratives. Hands stay on the keyboard.

Wearable

Plaud-style ambient pin.

Clip-on, all-day capture. A press-to-stop button when you step into a sensitive moment. Daily roll-up of dictations queued for review at 5pm.

In-room

Exam-room ambient mic.

Optional. A discreet, wired mic in select rooms for full-encounter ambient charting. Patient-consent prompt before each session, off by default.

Opt-in only

Every audio stream is transcribed on the rig — never uploaded to a third party. Patient names are detected and masked in any output that leaves the network. In-room capture requires a per-session consent prompt before recording starts.

What you get

Six jobs off your plate.

Each one removes a specific category of small task. Sequenced across four phases so value lands early and the team has time to absorb each piece before the next.

01

The clinical picture, before your first patient.

Labs that need sign-off, refills waiting, today’s schedule with flags. Delivered before the first visit. The day starts with everything pulled into one place.

Clinical · on-device
02

Prior auths drafted from the chart.

"Draft the Humana PA for Mrs. Kim." The agent pulls labs, history, and step-edit context, drafts the letter, and queues it for your signature. No portal hunting.

Clinical · on-device
03

Licenses and credentialing, never missed.

DEA, state license, malpractice, CME, hospital privileges, payer credentialing. All renewals tracked in the background. Alerts long before any deadline.

Business · BAA-free
04

Patient messages drafted in your voice.

Triage on every portal message. Five reply workflows trained on your real exchanges. Faster responses on routine questions, calmer ones on sensitive ones.

Clinical · on-device
05

The back-office runs while you see patients.

Staff scheduling, vendor follow-ups, supply orders, equipment maintenance. The administrative work the practice runs on, organized and watched in the background.

Business · BAA-free
06

Billing denials surfaced before they age.

The clearinghouse watched continuously. Denied claims flagged the same day with the reason and a suggested resubmission. Revenue that was quietly walking out the door, caught.

Clinical · on-devicePhase 3

Deliverables

What lands on your desk.

Three artifacts, in motion. Styling will be tuned to your voice and the practice’s brand as we build.

Conversation · Phase 1

Draft the Humana PA from the chart.

Plain-English requests turned into prior-auth letters. The Assistant pulls the chart, validates step therapy, drafts the justification, and queues the letter for your signature.

Practice Assistant · on-deviceEHR connected
Ask the Assistant…

Morning brief · Phase 1

7:30 AM, every day.

Labs to sign, refills pending, today’s schedule with flags. Pulled from the EHR overnight and delivered before the first patient. The day starts with everything in one place.

FromPractice Assistant · Cedar Family Medicine
SubjectMorning brief · May 19 · 12 patients, 3 labs, 7 refills
Monday · 7:30 AM

Good morning, Dr. Chen. Here’s the morning, in one read.

Labs to sign3 results
Susan Kim · A1c, lipid panelFlagResult4/30

A1c 7.8%. LDL up to 142. Up-titration consideration.

Marcus Chen · TSH, free T4Result5/01

TSH 6.4. Subclinical hypothyroidism on screening.

Refills pending7 · 2 controlled
Maria Garcia · sertraline 50mgRefill5/02

Last visit 2 mo ago. PHQ-9 unchanged. Standard renewal.

Today’s schedule12 patients · 1 new
Walter Park · new patient consultFlag10:30Rm 2

Insurance verified. Records from prior PCP not received.

Linda Okafor · follow-upReturning2:15Rm 1

Returning. Recent ER visit, records pulled and summarized.

Renewals · Phase 1

Nothing renews late.

DEA, state license, malpractice, CME, hospital privileges, payer credentialing — all watched in the background. Urgent items raised early. Routine items added to the queue. Done items kept on file.

Credentialing & renewalswatching 14
UrgentDEA registration · Dr. Chenexpires Jun 3 · renewal draftedJun 3
This weekMalpractice renewalawaiting carrier replyMay 24
RoutineCME · 10 hrs remaining3 credits logged this monthDec 31
RoutineHospital privileges · Dr. Parkrecredentialing packet openAug 15
DoneCLIA certificaterenewed · confirmedApr 15
DoneState medical licenserenewed · confirmedApr 15
0 of 14 renewals on track this quarter

How we work

Built around your practice.

The same rigor we’ve brought to large enterprise deployments, applied at the scale of a partner-led practice. Hands-on, founder-led, and end-to-end.

The AI runs in your building.

Models execute on hardware you own, on your network. Patient data is not transmitted to a cloud inference endpoint. No vendor BAA, no third-party processing.

Human approval on every write.

Every prior auth, patient reply, and clinical draft requires your explicit sign-off. The AI prepares, you decide. Nothing leaves under your name without you.

Voice trained on your letters.

Calibrated on 15–20 of your real referral letters and PA templates before anything goes out. The agent writes in the cadence and clinical tone you already use.

Air-gappable by design.

The rig can run fully offline. Cloud is opt-in for specific non-PHI workflows (Google reviews, public web). The default posture is local, encrypted at rest, and yours.

How we get there

The build, in four phases.

Phases 1 and 2 are in scope for the current Statement of Work. Phases 3 and 4 are mapped here so the partners can see the full arc, with value landing early and risk staying low.

Phase 1Weeks 1–4

The Foundation

Rig deployed, morning brief landing, PA assistant live

We deploy the rig, connect the EHR, and stand up the on-device model layer. The morning clinical brief lands daily. The PA drafting assistant is ready for its first letter. Credentialing tracker populated.

  • On-site AI rig deployed and configured
  • EHR & lab inbox connected, models running locally
  • Morning clinical brief, by email
  • Prior auth drafting assistant
  • Credentialing & renewals tracker
Phase 2Weeks 5–9

The Inbox & Voice

Patient messages, voice capture, back-office

The patient inbox triage agent goes live. Voice capture rolls out across mobile, tablet, laptop, wearable, and exam-room mics. The back-office layer (staff, vendors, supplies) comes online. A draft-only week for calibration, then approval-and-send.

  • Voice training on your real exchanges
  • Five drafted-reply workflows, daily
  • Voice capture across 5 surfaces
  • Staff, vendor, and supply back-office
Phase 3Weeks 10–13Planning

Billing Watch

Denials surfaced before they age out

The clearinghouse runs in the background. Denied claims flagged the same day. AR aging digest delivered weekly with proposed resubmission paths. Revenue captured that was quietly being written off.

  • Continuous clearinghouse monitoring
  • Same-day denial flagging
  • Resubmission letters drafted from EOBs
  • Weekly AR aging digest
Phase 4Month 4+Planning

Practice Growth

Recall, reviews, new patient flow

Care gap outreach for chronic disease panels. New patient follow-up. Google review triage and response drafts. Schedule analytics so the practice runs at the rhythm you want.

  • Care gap outreach (A1c, lipid, BP)
  • New patient pipeline & follow-up
  • Google review triage and drafted responses
  • Schedule analytics & capacity planning

About

Who’s building this.

River AI is Brian Joseff and Chandler Supple. We design and deploy custom AI operations layers for businesses where the work is too specific for generic software. The same depth we’ve brought to Fortune 100 environments, scaled to fit a single physician-led practice.

We work hands-on, founder-led, and end-to-end. You aren’t handed off to an account team. You build directly with us, and the people who design the system are the people who ship it.

Brian JoseffCo-founder
Chandler SuppleCo-founder

Next steps

Three steps. Then Week 1.

Nothing for you to schedule. The Statement of Work is drafted from our conversation and ready in your inbox. Review it with the partners and we’re underway.

01

The Statement of Work is in your inbox.

A tailored scope, a fixed price, and a clear definition of done for Phases 1 and 2. Phases 3 and 4 are mapped for planning. Drafted from our conversation, ready for the partners.

02

You review and sign with the team.

We answer any questions, adjust scope if anything is off, and only kick off once the SOW is comfortable for everyone involved. Signature triggers a 25% deposit.

03

Week 1 begins.

We provision the rig, ship it on-site, and authenticate the EHR. The first morning brief lands at the end of Phase 1. The first piece of evidence in your inbox.