How we make this possible

One visit. Two doctors. One record the patient can read.

A visit in Spanish with a physician in Latin America. A prescription from a physician licensed in the patient's state. Between them, the patient's information has to travel safely, mean the same thing on both ends, and come back in words the patient can act on. That is the job of CodeWeave, our medical data architecture, and the AI layer that runs on it.

01 · The IPCodeWeaveProprietary storage that codifies each clinical value and keeps the context in which it was collected and used. That context is what makes the record useful to LLMs, and it gives other systems a standard shape to read.
02 · The workloadSeven LLM tasks per patient journeyFour on the clinical side: transcription, translation, extraction and coding, provider documentation. Three on the patient side: literacy inference, guided conversations, plain-language explanations. Every output exists in Spanish and English.
03 · Cost-stable by designThe 10,000th visit costs what the first didPer-visit inference does not grow with the size of a patient's record. Total workload scales with patients and visits, linearly and predictably, which makes it easy to forecast.

The problem we solve

Health literacy is a question of availability

Not availability in the sense that the data can be accessed, but in the sense that health information is consumable and understandable by every stakeholder in a patient's journey. Most healthcare data platforms forget the person to whom the data matters most: the patient.

Azam is built to fix that. CodeWeave makes the data usable by machines and by other systems. The AI layer on top of it makes the same data usable by physicians in two countries and, above all, by the patient, in their language and at their level.

Architecture

From a conversation to a coded record, and back to people

LLM task
Proprietary data layer
People
Sign-offA licensed human approves before anything is final

Capture · per visit

Patient appguided conversation, ES / EN

Video visitLATAM physician + patient

Transcriptionvoice to text, speaker ID

Translationacross languages and locales

Extraction + codingclinical facts to SNOMED CT

CodeWeaveEach clinical value stored with the context in which it was collected and used: source, speaker, language and locale, setting, time, confidence. Codified for machines, contextual for LLMs, standard for other systems.

Outputs

Provider documentationin the physician's own voiceSign-off

U.S. prescriber handoffphysician licensed in the patient's stateSign-off

Health-literacy modelinfers the patient's level from their own interactions

Patient-facing layerguided forms and plain explanations, at the patient's level

A visit is captured in five steps: the patient app's guided conversation and the video visit itself involve people; transcription, translation, and extraction to SNOMED CT codes are model tasks. All five write into CodeWeave. Four things read back out of it: the provider's documentation and the U.S. prescriber handoff, both of which a licensed physician signs, and the health-literacy model, which feeds the patient-facing layer.

CodeWeave

The value, and the context that makes it mean something

Most health data stores keep a value: "fever: yes". CodeWeave keeps the value and the qualifying information around it. That is what lets a model, or a second physician, tell a symptom the patient reported at intake from a finding the physician observed on video.

Codified values.Every clinical fact is mapped to a standard concept (SNOMED CT), so it can be counted, compared and queried.
Context of collection and use.Who said it, in which language and locale, in what setting, when, with what confidence, and what it was later used for.
Interoperable by design.The same shape is what the U.S. prescribing physician reads, and what an external system would receive.
Constant cost to read.A visit is processed against a bounded, structured record, not a growing transcript history. Per-visit inference stays flat as the record grows.

One CodeWeave entry · illustrative

{
  "concept": "SNOMED CT 386661006",   // fever
  "value": "present",
  "qualifier": "nocturnal",
  "context": {
    "source": "patient-reported",
    "speaker": "patient",
    "utterance_lang": "es-SV",
    "surface_form": "calentura",
    "setting": "guided intake",
    "captured_at": "[timestamp]",
    "confidence": "[score]",
    "used_by": ["provider-note", "rx-handoff", "patient-summary"]
  }
}

The English rendering, the SNOMED code and the original Salvadoran word all live together. Nothing is lost in translation, because the translation is stored next to the source.

Worked example · illustrative

One visit, start to finish, in both languages

A patient in Chicago, originally from El Salvador, has a sore throat. The dialogue is illustrative and contains no patient data.

1 · Guided intake
Spanish (as spoken)
App: «Cuéntame, ¿qué te está pasando?» — «Me duele la garganta desde el lunes y en la noche me da calentura.»
English rendering
App: "Tell me, what's going on?" — "My throat has hurt since Monday and at night I get a fever."
What the system produces
The standard intake form (chief complaint, onset, associated symptoms) is asked as a conversation. "Since Monday" resolves to a date. The regional word calentura is mapped to fever, and kept.
2 · Visit transcription
Spanish (as spoken)
[DR] ¿Ha tenido tos? · [PT] Poquita, seca.
English rendering
[DR] Have you had a cough? · [PT] A little, dry.
What the system produces
Timestamped segments, each labeled with who spoke. Kept in both languages.
3 · Extraction + coding
Spanish (as spoken)
Dolor de garganta · Calentura nocturna · Tos seca leve
English rendering
Pain in throat · SNOMED CT 162397003 · onset [date] Fever · SNOMED CT 386661006 · nocturnal Dry cough · SNOMED CT 11833005 · mild
What the system produces
Structured findings, each linked back to the exact words it came from, and stored as CodeWeave entries.
4 · Provider noteSign-off
Spanish (as spoken)
Nota S/O/A/P redactada en el estilo del médico, revisada y firmada por él.
English rendering
Draft S/O/A/P note generated in the LATAM physician's own documentation style, with co-pilot suggestions for the assessment. The physician edits and signs it.
What the system produces
Compliance documentation, in the provider's voice, in both languages. Nothing enters the record without the physician's signature.
5 · U.S. handoffSign-off
Spanish (as spoken)
El expediente firmado y codificado pasa a un médico con licencia en Illinois.
English rendering
If a prescription is indicated, the signed, coded record goes to a partner physician licensed in Illinois, who reviews it and issues, adjusts or declines the prescription.
What the system produces
The U.S. physician reads the same CodeWeave record the LATAM physician signed: no re-telling, no second intake.
6 · Back to the patient
Spanish (as spoken)
«Tu doctora cree que es una infección de garganta. Un médico de Illinois revisó tu caso y te mandó una medicina. Tómala 3 veces al día con comida, por 7 días. Si la calentura sigue después de 3 días, escríbele aquí.»
English rendering
"Your doctor thinks it's a throat infection. A doctor in Illinois reviewed your case and sent you a medicine. Take it 3 times a day with food, for 7 days. If the fever continues after 3 days, message her here."
What the system produces
The explanation restates what the two physicians decided, at the patient's inferred reading level, using the patient's own word for fever. It adds no medical advice of its own.

The health-literacy layer

Information the patient can understand and own

Transcribing, translating and coding a visit is a step forward for the data. On its own it does nothing for the patient's understanding. This layer is what does, and it is the part of the system we are proudest of.

1 · InferEstimate the patient's health literacyAn LLM analyzes the patient's own interactions, such as visit transcripts and messages, to identify their health-literacy level. No quiz, no label the patient has to pick. The estimate updates as they use the app.
2 · AskTurn forms into guided conversationsStandard medical forms and questionnaires are converted into a conversation, using the literacy level, the patient's preferred language and their cultural context. The clinical fields still get filled. The patient never sees the form.
3 · ExplainPresent critical information in a voice the patient can act onVisit summaries, prescriptions, lab results and translated documents come back understandable, consumable and actionable. The patient can ask "why am I taking this?" and get an answer at their level, grounded in what their physician decided.

A form field, before and after · illustrative

Standard form: Antecedentes familiares: enfermedad cardiovascular ☐ Sí ☐ No
Guided conversation: «¿Alguien cercano en tu familia, como tus papás o hermanos, ha tenido problemas del corazón?»

A result, before and after · illustrative

In the record: Hypertension, stage 1 · SNOMED CT 38341003 · physician plan: lifestyle, recheck 4 wk
To the patient: «Tu presión está un poco alta. No es una emergencia. Tu doctora quiere que comas con menos sal y camines más, y volver a medirla en un mes.»

Privacy and the U.S. physician network

Why the record has to travel, and how it travels safely

The clinical visit happens in Latin America. Prescribing has to happen with a physician licensed in the patient's state. CodeWeave exists so that handoff can happen without the patient repeating themselves, and without their information leaving their control.

The visit ends and the LATAM physician signs the note.The CodeWeave record for that visit is complete: coded findings, the physician's assessment and plan, in both languages.
If a prescription is indicated, the signed record is routed to a partner physician licensed in the patient's state.They receive the structured, coded case, not a recording and not a chat log.
The U.S. physician reviews and decides.They issue, adjust or decline the prescription under their own license. Their decision is recorded in the same record.
The patient sees the outcome, explained.The prescription, who issued it, and what to do next, in plain language in the app. The patient can read their full record at any time.

Patient rights and privacy

Records are stored and shared in line with U.S. health-privacy rules (HIPAA). [Legal review: BAAs with model providers, data residency, retention]
Only the physicians treating the patient see their record.
Patients can read their own record, in their language, at any time.
No patient data is sold. Model providers do not train on it. [Confirm contractually]

The U.S. partner network

Physicians licensed in the states we serve, starting with Illinois, Texas and California. They review encoded cases asynchronously and prescribe under their own license.
[Number of partner physicians, states and the prescribing model to confirm]

Inference workload

Flat per visit. Linear across patients.

We designed the pipeline so that the cost of processing a visit does not depend on how much history the patient has. That makes the workload easy to forecast: it is a function of how many people we serve and how often they use the service.

[__]minutes of audio transcribed per visit
[__]model calls per visit
[__]ktokens per visit, constant
2languages for every artifact
7LLM tasks across the patient journey

Per patient, per month

tokens / patient / month =
    visits    × C_visit
  + messages  × C_message
  + documents × C_document

C_* are constants. None of them depends
on the size of the patient's record.

Each visit runs against a bounded, structured CodeWeave view, not the full history. The first visit and the ten-thousandth cost the same to process.

Across the platform

Scenario	Active patients	Tokens / month
Today, El Salvador	10,000+	[__]
U.S. launch: Chicago, Houston, LA	[__]	[__]
12 months out	[__]	[__]

Volume grows with patients and with visits per patient, and every output is produced in two languages. Because per-event cost is constant, the forecast is a multiplication, not a curve.

Safety posture

The model helps. Clinicians decide.

Physicians sign offThe co-pilot suggests. The physician edits and signs every note. Nothing enters the chart without a physician's signature.
The platform doesn't give medical advicePatient explanations restate the physicians' plan in plainer words. They don't diagnose or recommend treatment on their own.
U.S.-licensed prescribingEvery prescription is issued by a physician licensed in the patient's own state, under their own license.
Data protection[Pending legal review: HIPAA posture, BAAs with model providers, data residency, retention, encryption at rest and in transit]
Evaluation[CTO to supply: how we measure transcription, translation, coding and literacy-adaptation accuracy, and how physicians flag errors]
Emergencies stay out of scopeAzam is not for emergencies. Patients are told to call 911, in every language and at every literacy level.

Team

Alejandro Siman · CEOFounder. From El Salvador; Notre Dame '28.
JP Burford · CTOMulti-time CTO, Y Combinator alum, prior healthcare exit. Architect of CodeWeave.
Dr. David Gaus · Medical AdvisorFounder of Andean Health.

A growing community

10,000+

people are part of Azam in El Salvador, where the service has run for over a year in partnership with MAPFRE's medical network

First U.S. markets: Chicago, Houston, Los Angeles.

Questions about the architecture?

We'll walk you through the pipeline and share workload data.

alajandro@azamhealth.app