Atlas & AI
Atlas is Ready Practice's AI: staff copilot + clinical assistant + consumer Health Assistant + COO/operator + Slack employee — one identity across surfaces, backed by the agent fleet.
The surfaces of Atlas
- Copilot (staff) — in-app assistant, doc/lab analysis, drafting. → copilot
- Consumer Health Assistant (patients) — text-only, approval cards, protocol-apply is clinician-gated. → atlas-consumer-handoff, chat-atlas-spec
- Operator / COO — autonomous operations. → atlas-operator-plan
- Slack employee — inbound ops Q&A, approvals as buttons, earned attention pings. → atlas-slack-plan
- Deal-runner — one engine, three counterparty types (sales / clinician superconnector / fundraise). → atlas-dealrunner-roadmap
Model tiers
Never hardcode model IDs — import from atlas_models.py:
| Tier | Model | Use |
|---|---|---|
ATLAS_FAST | Haiku | Structured / clear tasks |
ATLAS_BALANCED | Sonnet | Long tasks |
ATLAS_SMART | Opus | Conversational / hardest reasoning |
functions_model_watch.py auto-detects new/deprecated models.
Guardrails
- Untrusted text → LLM prompts must be fenced (
fence_untrusted()/ explicit===== UNTRUSTED DATA =====) — any error/user/issue-sourced string is a prompt-injection vector. Linted at predeploy + CI. - User opt-in required (
preferences.llmOptIn), HIPAA-aware prompts, no training on patient data. - Adding a tool touches 5 silent-fail layers — see Engineering.
Where this is heading
The retrieval + orchestration roadmap (hybrid retrieval, parallel fan-out, adversarial verification, per-clinic PHI-isolated KB) is written up in Knowledge Base Vision.
Build state
Snapshot: atlas-build-state. The AI voice /
personality is defined in basis-functions/functions/src/prompts/consumer/soul.md —
see Brand & Voice.