Alyssa

Conversational API endpoint that generates SMS and voice responses for MD Well patient interactions. Two workflows: booking first coaching calls (sales) and weekly check-ins with enrolled patients (care). Every response passes inline quality gates (HIPAA, TCPA, crisis detection) before it reaches your system.

You send a patient message and conversation history. Alyssa returns the response. Your system handles delivery, scheduling, and escalation routing.

Send a patient message, get Alyssa's response:

curl -X POST https://{instance}.culturedcomputer.com/api/chat \
  -H "Authorization: Bearer {api_key}" \
  -H "Content-Type: application/json" \
  -d '{
    "message": "How much does this cost?",
    "userId": "patient-mike-001",
    "model": "mercury",
    "toneName": "alyssa",
    "avatarName": "Alyssa",
    "conversationHistory": [
      { "role": "assistant", "content": "Hi Mike, it'\''s Alyssa from MD Well" }
    ],
    "templateParams": {
      "first_name": "Mike",
      "doctor_name": "Chen",
      "clinic_name": "MD Well",
      "booking_link": "https://mdwell.com/book"
    }
  }'

Response (Server-Sent Events):

data: {"choices":[{"delta":{"content":"Good news"}}]}
data: {"choices":[{"delta":{"content":", it's covered"}}]}
data: {"choices":[{"delta":{"content":" by your insurance."}}]}
data: {"choices":[{"delta":{"content":" Want to set up"}}]}
data: {"choices":[{"delta":{"content":" a quick call to get started?"}}]}
data: [DONE]

802ms to first character, 861ms to completion.

Alyssa follows the MD Well outreach SOP. Every response uses the Younes framing and the patient's doctor name from templateParams.

Objection Handling

Production output

Alyssa adapts based on the patient message and conversation history you provide.

Response Patterns

Production output

Two-call pattern: a short acknowledgment (call 1) streams to TTS immediately so the patient hears something fast, then a full response (call 2) follows.

The API generates responses. Your system handles everything else.

Conversation History

Store each exchange and pass the full array back on every /chat call. Alyssa uses this for continuity, tone adaptation, and personalization.

"conversationHistory": [
  { "role": "assistant", "content": "Hi Mike, it's Alyssa from MD Well" },
  { "role": "user", "content": "who is this?" },
  { "role": "assistant", "content": "Dr. Younes set up a coaching program..." }
]

Patient Context

Provide templateParams from your patient records on every call. Alyssa resolves these into responses automatically.

Follow-Up Scheduling (sales)

When a patient does not reply, your scheduler triggers follow-up calls to /chat. The API generates the content; you control the timing.

If the API returns no response (TCPA trigger), stop the sequence.

Check-In Cadence (care)

Your system initiates check-ins by calling /chat. Alyssa adapts her tone based on conversation history.

Alyssa's coaching focus evolves over a 12-week arc based on conversation depth.

Escalation Routing

When the API returns escalation.required: true, route the conversation to a human coach or clinician. Alyssa does not generate coaching content for these messages.

Opt-Out Management

When the API returns no response (patient sent STOP, unsubscribe, or scam), mark the patient as opted out in your system and stop sending.

Base URL: https://{instance}.culturedcomputer.com/api
Authentication: Authorization: Bearer {api_key}

POST/chat

Send a patient message. Returns Alyssa's response as Server-Sent Events.

Response: SSE stream. Each chunk: data: {"choices":[{"delta":{"content":"..."}}]}. Final: data: [DONE].

POST/eval-explain

Score a response against quality gates.

GET/model-status?model=mercury

Returns model availability and current latency.

GET/qa-latency?mode=production

Runs all test scenarios (3 care + 4 sales + 1 voice) with full quality gates.

POST/session/save

Persists a completed conversation with metadata: turns, duration, quality scores, escalation flags.

Every response is checked inline. Tier 1 failures block the response before it reaches your system.

Pass Rates (production, 7 SMS scenarios)

32K token context window per conversation (~20 turns).

Everything above works today. This section describes the next iteration: richer patient context, conversation state tracking, and goal-oriented care workflows. These features use infrastructure that already exists in the platform.

Extended Patient Context

The current templateParams accepts 5 fields. v2 expands this so Alyssa can generate track-specific, condition-aware, arc-appropriate responses.

"templateParams": {
  "first_name": "Mike",
  "doctor_name": "Chen",
  "clinic_name": "MD Well",
  "booking_link": "https://mdwell.com/book",
  "coach_name": "Alyssa",

  "track": "ccm",
  "conditions": ["type_2_diabetes", "hypertension"],
  "goals": ["lower A1C", "walk 10 min/day"],
  "chosen_microhabit": "10-minute walk after dinner",
  "medications": ["Metformin", "Lisinopril"],
  "insurance_status": "covered",
  "last_visit_date": "2026-02-15"
}

Conversation State

v2 adds conversationContext to each call. This lets Alyssa track engagement depth and adapt behavior without your system needing to interpret conversation content.

"conversationContext": {
  "eval": {
    "state": "ENGAGED",
    "depth": 4
  }
}

Goal-Oriented Workflows

v2 adds goalId to steer Alyssa toward specific care objectives. Each goal defines success signals, failure signals, and state-specific behavior.

What This Enables

Your system remains the orchestrator: scheduling, delivery, cadence, opt-out management, clinical routing. The API becomes a richer conversational engine that understands where each patient is in their care journey.