Investor Brief · Part 2
A pragmatic stack that's standards-based at the bottom, frontier-LLM at the reasoning layer, and trace-anchored throughout. Each layer below is tagged with current status (Shipped / Wired in prototype / Designed / Future) and cross-links to the architecture brief that owns it.
Stack overview
FHIR R5 via JupyterHealth Exchange (primary), HL7v2 fallback, MuleSoft API-Led Connectivity for system integrations.
Standards-based. JupyterHealth is the consented FHIR substrate; MuleSoft handles system-to-system writes and transforms. Avoids per-EHR custom work.
JupyterHealth integration brief →Salesforce Data 360 — Identity Resolution + Calculated Insights + Segments, zero-copy federated over JupyterHealth FHIR, DBDP features, and the customer's EHR-of-record.
Solves identity + grounding in one substrate without bulk PHI ingestion. Snowflake / Databricks / Epic Health Cloud are federation targets, not separate Pause platforms.
Data 360 brief →Anthropic Claude Sonnet 4.5 via @anthropic-ai/sdk for the Care Router; deterministic Pause policy engine as fallback. Multi-agent control plane via the (mocked) MuleSoft Agent Fabric.
Frontier-quality reasoning where it counts; deterministic safety floor where it doesn't. The Agent Fabric provides registry + policy catalog + trace plane.
Agent Fabric brief →MCP server exposing four Pause tools (get_patient_timeline, get_data360_grounding, evaluate_routing_policy, list_segments). Compatible with Claude Desktop, Cursor, Agentforce, and any MCP client.
Lets the customer's own LLMs read Pause's data + policies the same way our agents do. Standard protocol, no per-vendor adapters.
MCP server brief →Frontier providers (Anthropic primary, OpenAI / Google as alternates). Open-weight options (Llama, Mistral) earmarked for security-sensitive on-prem deployments.
Best-in-class quality where allowed; on-prem swap available for customers that need it, without rebuilding the agent layer.
Phase 2: domain-specific classifiers and embeddings for symptom-cluster scoring + risk stratification. Phase 1 uses the deterministic policy engine + LLM reasoning.
Small models for the repetitive structured work; LLMs only where reasoning is required. Specialty models are a Phase 2 step gated on having labeled real-deployment data.
Digital Biomarker Discovery Pipeline (DBDP). Phase 1 shipped: FLIRT-backed RMSSD with closed-form correctness tests. Phase 2: EDA, sleep, vasomotor burden composite via Devicely + DHDR.
Reuses peer-reviewed wearable feature pipelines instead of reinventing them. Honest about what's shipped (RMSSD) vs what awaits library upgrades (numpy < 2.0 in Devicely's current release).
DBDP feature engineering brief →Next.js (this site, the SMART-on-FHIR launch surface ahead). Salesforce Agentforce Service Agent for live intake (running on a real Service Cloud org today).
Same engineers can move between web, EHR-embedded, and intake surfaces. React Native patient-mobile app is Phase 2.
Agentforce intake brief →Vercel for the demo + investor brief; AWS (HIPAA-eligible) + customer-controlled VPC for production deployments. GitHub Actions CI.
Vercel today because the brief is a public artifact; AWS + per-customer VPC at customer-deployment time so PHI never leaves customer-controlled blast radius.
Agent Fabric trace ring buffer: per-call spans with model, inputs, outputs, sources queried, durations, agent + protocol. Customer-deployment OpenTelemetry export via the MuleSoft tap.
Audit-grade trail. Inputs/outputs/model/duration are live today in the demo trace plane; clinician-action capture lands when the SMART-on-FHIR install ships.
See the live trace plane →AI approach
Anthropic Claude (Sonnet 4.5) prompted with the patient's Data 360 grounding payload + a guideline-aware system prompt. Outputs the canonical pathway enum (six values) + acuity + rationale, with the deterministic policy engine as fallback when the LLM is unavailable.
Phase 2 work. Today's prompts are guideline-aware but not retrieval-grounded against a corpus. Phase 2 builds the curated menopause guideline corpus + retrieval pipeline + citation surfacing.
Phase 1 today: deterministic symptom-cluster scoring via the policy engine + LLM-assisted free-text interpretation. Phase 2: domain-specific embedding model trained on PRO + EHR free text once we have labeled deployment data.
Phase 1 today: deterministic risk band from intake scores (V+S+M aggregate + per-axis flags). Phase 2: gradient-boosted classifier on structured features once labeled deployment data exists.
Provider mode: clinical pre-read + decision support inside the Care Detail view (already live). Patient mode: live Agentforce-driven intake (already live). Roles enforced by the Agent Fabric policy catalog.
Today: every Care Router decision lands in the trace plane with inputs, outputs, model version, and duration. Tomorrow: clinician edits/rejections feed preference-tuning datasets + the outcomes registry validates long-term accuracy.
Evaluation framework
Every target below is a measurement plan, not a current observation. Pause is pre-design-partner; the eval framework becomes a measurement contract with the first IDN.
| Metric | Target | Why it matters |
|---|---|---|
| Clinician acceptance rate | >= 70% accept-or-edit; < 10% reject outright | Operationally meaningful adoption signal; precedes outcomes data. |
| Recommendation accuracy vs. specialist panel | >= 85% concordance on top-1; >= 95% on top-3 | Validates that the system suggests what an MSCP-credentialed clinician would. |
| Diagnostic time reduction | From 2.5 years (industry average) to < 90 days for newly-onset cases | Direct patient outcome and a compelling marketing claim — only credible if measured. |
| Avoidable utilization reduction | 10–20% reduction in ER + specialist visits for the cohort over 12 months | Anchors the payer ROI conversation. |
| Hallucination rate on guideline questions | < 1% on a held-out evaluation set; 0% on contraindication questions | Safety floor. Contraindications + dosing must be deterministic, enforced by the Agent Fabric policy catalog. |
Safety and trust
The system addresses menopause-related decisions only. Out-of-scope questions are deflected with an explicit handoff. Enforced by the agent's system prompt + the Agent Fabric scope policy.
No autonomous prescribing. No autonomous patient messaging without clinician review. The 'no-clinical-autonomy' policy in the Agent Fabric policy catalog enforces this on every span.
Every Care Router decision carries inputs, model, rationale, and sources-queried as span attributes. The trace plane surfaces them; the Care Detail view shows the same to clinicians.
Sub-group performance tracking by age band, race, ethnicity, geography, and clinical setting — reported quarterly to the clinical advisory board. Pipeline + dashboards are Phase 2 work gated on deployment data.
Quarterly red-team exercises by external clinical evaluators. Findings published internally and acted on before each major release. Process is documented; first formal red-team is gated on first design partner.
HIPAA-ready architecture today (zero-copy federation, no bulk PHI ingest, customer-controlled blast radius in production). HITRUST CSF + SOC 2 Type II are on the roadmap.
Read deeper
Each of the stack layers above cross-links to its own brief. Following the stack from interop floor to control plane:
Continuing the investor brief
