Skip to content
AskFlorence
Main Navigation ArchitectureFlorence AIAgentsMembersAgent PlatformValidationInfrastructure

Appearance

Sidebar Navigation

Overview

Home

Glossary

System Architecture

Consumer & Agent Flow

Florence AI

Overview

Principles

Runtime

Tool surface

Adding a tool

Tool registry

Knowledge: SBC scenarios & CSR

Voice

Evals & observability

Provider risk & portability

Outage playbook

Roadmap

Build plan

Agents

Overview

Workflows & pain points

Members

Overview

Medicaid coverage gap

Carriers

Overview

Marketplaces

Overview

Agency

Overview

Regulations

Overview

Agent Platform

Overview

Auth Architecture

MongoDB Permissioning

Compliance Model

Data Models

Data Sources

Overview

CMS Marketplace API

CMS dependency map

PUF Data

State Subsidies

SBE Ingestion Playbook

SBE State Watchouts + Decisions

CA Phase C/D Playbook

NY Phase C/D Playbook

Validation

Overview

Methodology

APTC Formula

California 2026

New York 2026

CAPS Formula

Scenario Results

Infrastructure

Account Inventory

AWS Setup Runbook

AWS Organizations

CloudTrail

GuardDuty

Security Hub

Config

CloudFront + WAFv2

Data sources & ingest

Phase 4 DNS

Change Log

Vulnerability Management

MongoDB Setup

Access Control

Data Classification

Documentation Hosting

Post-deploy Smoke

Development

Preflight (local CI mirror)

Testing strategy

Compliance

Overview (auditor entry point)

SOC 2 Control Mapping

HIPAA Control Mapping

CMS EDE Appendix A Mapping

Risk Assessment

Encryption Policy

Data Retention Policy

Privacy Impact Assessment

Consent Capture & Versioning

Incident Response Plan

Access Control Policy

Marketing vs. Portal Analytics

Vendor / Subprocessor Register

Dependency Vulnerability Policy

BAA / Compliance Evidence

Compliance-Automation Integration

Compliance-Automation Vendor Evaluation

Penetration Test Reports

Architecture

Portal entry handoff

Mobile app strategy

Deferred architecture decisions

Session cookie architecture

Share flows

Decisions (ADRs)

Index

0001 — Atlas project isolation

0002 — Append-only audit log

0003 — Narrow-scoped Mongo users

0004 — Cross-cluster Atlas PrivateLink

0005 — Delayed-job architecture

0006 — Mongo user simplification

0007 — Terraform owns ECS task def

0008 — E2E testing strategy

0009 — Self-hosted analytics + observability (superseded)

0010 — PostHog HIPAA Cloud (supersedes 0009)

Runbooks

Security Incident Response

Break-Glass Root Login

Onboard Team Member

Offboard Team Member

Atlas user provisioning

Deploy via Terraform (ENG-277)

Rollback via Terraform (ENG-277)

S3 data bucket migration (planned Phase 11)

Access Reviews

2026-Q2 Review

Session log

Index

2026-04-23 — Phase 10 DNS cutover

2026-04-22 — Phase 8 prod AWS mirror

2026-04-22 — Phase 7 Atlas VPC peering

2026-04-22 — Phase 6 CloudFront + WAF

2026-04-21 — Phase 5 staging go-live

2026-04-17 — Atlas staging

Briefs

Index

Member portal plan (ENG-187)

2026-04-16/17 handoff

2026-04-17 Atlas handoff

System briefing (2026-04-17)

Creative AdBundance proposal brief

Creative AdBundance analytics brief

ElevenLabs RN integration research

Policies

Overview

On this page

Florence AI — architecture ​

Status: Living document. v0.1, 2026-04-22. Tracking issue: #61 — Florence AI architecture + infrastructure design. Input brief (input-of-record): /briefs/BRIEF_FLORENCE_AI_ARCHITECTURE.

Naming note: the brand "Florence" has trademark exposure. In all new work use Florence AI or AskFlorence AI until legal clears the final brand. This directory is the settled-architecture home; the input brief above is preserved verbatim for fidelity to the original hand-off document.

Florence AI is the universal conversational interface to AskFlorence. Every user interaction — pre-enrollment plan discovery, intake, enrollment hand-holding, post-enrollment support, renewal, appointment assistance — flows through Florence.

The UI (plan cards, filters, forms, dashboards) and Florence are two faces of the same state machine: both drive the same deterministic application functions. Intercom-style support tools cannot exist in our Phase 3 EDE posture; Florence absorbs that role by design, not as an add-on.

The one sentence that governs everything ​

Florence is a natural-language interface over the deterministic AskFlorence platform. She never computes a fact. Every factual claim she makes traces to a tool call in the current turn.

That sentence is the north star. Every downstream decision in this directory derives from it.

How to navigate this directory ​

DocWhat it covers
PrinciplesThe invariants. Deterministic grounding, no-compute, text-as-source-of-truth, unit-economics targets, camouflage posture. Read first.
RuntimeFlorenceRuntime on Claude Agent SDK. Model routing, prompt caching, streaming, state persistence, context compression.
Tool surfaceHow tools work. api_* vs ui_* families. Auth context. The deterministic-integration pattern every tool follows.
Adding a toolPlaybook. Lifecycle from proposal → schema → classification → eval coverage → ship → monitor → deprecate. Security review required at each stage.
Tool registryLiving inventory. Every Florence-callable tool, its backing deterministic endpoint, data classification, auth scope, eval coverage, status.
Knowledge: SBC scenarios & CSRTool-callable reference for the plan-detail Year Scenarios block: CMS standardized totals, where puf.csrVariants[tier] lives in MongoDB, why CSR-94 patient cost can still be meaningful (AV ≠ per-scenario), dataset distribution, worked walkthrough Florence can quote, ENG-365 Option C provenance.
VoiceThree-stream architecture. Vendor strategy through Phase 3. EN + ES from day one. Quality-preserving paths to EDE.
Evals & observabilityEval harness design. Audit log + cost dashboard + alert policy. How we hold the unit-economics targets in practice.
Provider risk & portabilityVendor concentration risk, the four tiers of switch (Tier 0 transport → Tier 1 model → Tier 2 cross-vendor → Tier 3 self-hosted), capability matrix, warm-standby operational plan.
Outage playbookOperational companion to provider-risk. Failure-domain analysis, the 6-tier quality-preserving cascade, circuit breakers + thresholds, request hedging, Tier 6 deterministic-only mode, chaos drills. Validated against recent Anthropic incident history.
RoadmapPhase 1 → 1.5 → 3 sequencing. Dependencies on AWS migration (#47), compliance work (#55, #56, #57), data-classification rollout.
Build planConcrete staged build: A0 spike → A1 foundation → A2 beta on staging. Stream A (runtime), Stream B (tool PRs), Stream C (data-classification retrofit), Stream D (infra). Parallel-safe branch discipline + handoff prompt for the first Stream A session.
WOW flow researchENG-356. Clicky teardown (incl. their ElevenLabs wiring + rendering craft), Cartesia vs ElevenLabs eval with the demo-acceptable vs production-BAA-signable distinction + transcript-as-PHI, the proposed voice-synced rendered WOW flow (beats x scene x tool x spotlight + state machine + edges), demo-vs-prod brain wiring, and the local demo.

What this directory is NOT ​

  • Not the deterministic platform's architecture (see system architecture and data sources).
  • Not the agent portal (see agent platform).
  • Not the infrastructure runbooks (see infrastructure).
  • Not an input brief. Input briefs live in /briefs/; this directory holds the settled architecture they informed.

Relationship to the rest of the system ​

The UI and Florence operate on the same deterministic platform. Florence can drive the UI via ui_* tools (filter a search, open a plan, highlight a form field). This is the co-pilot register that separates Florence from a chat widget.

Pager
Previous pageConsumer & Agent Flow
Next pagePrinciples

AskFlorence Internal Documentation. Not for public distribution.

AskFlorence

Internal Documentation

Access restricted. Not for public distribution.