Integration / ingestion
FHIR R4, SMART-on-FHIR, HL7v2 fallback, HealthKit / Health Connect bridges, X12 claims
Standards-based. Avoids per-EHR custom work. Lets us land in any Epic / Cerner site within weeks instead of quarters.
Investor Brief · Part 2
Technology Choices: architecture, AI approach, and safety
A pragmatic stack that earns clinician trust on day one and gets faster, sharper, and safer with every deployment.
Stack overview
Data platform
Snowflake / Databricks (per-customer choice), dbt for transforms, OMOP CDM alignment
OMOP gives us research-grade interoperability with academic partners and lets us reuse open methods.
Inference and orchestration
LangGraph-style stateful agents, retrieval-augmented generation over the guideline corpus, structured tool-calling
Modular, testable, and explainable. Every step of a recommendation can be replayed deterministically.
Foundation models
Frontier providers (Anthropic, OpenAI, Google) plus open-weight options (Llama, Mistral) for federated deployments
Best-in-class quality where allowed; on-prem option for security-sensitive customers without rebuilding the product.
Specialty models
Fine-tuned classifiers for symptom-cluster scoring, risk stratification, and PRO interpretation
Small, fast, evaluable models for the repetitive structured work; LLMs only where reasoning is required.
Application
Next.js (SMART-on-FHIR launch), React Native (patient mobile), TypeScript throughout
Same engineers can move between web, EHR-embedded, and mobile surfaces. Reduces silos.
Infrastructure
AWS (HIPAA-eligible), single-tenant VPC per enterprise customer, GitHub Actions CI
Health-IT-ready posture from day one. Customer-controlled blast radius.
Observability
Per-recommendation traces (inputs, retrieval, model version, output, clinician action)
Audit-grade trail. Required for clinical trust and for the training-feedback loop.
AI approach
Recommendation generation
RAG over a curated menopause guideline corpus + patient timeline. Outputs are structured, ranked, and accompanied by retrieved evidence.
Symptom clustering
Domain-specific embedding model trained on PRO + EHR free text. Identifies multi-system menopause presentations that single-symptom checklists miss.
Risk stratification
Gradient-boosted classifier on structured features (vitals, labs, history, wearable signals). Outputs are calibrated probabilities, not opaque scores.
Conversational interface
Constrained, role-aware agent. Provider mode: clinical pre-read and decision support. Patient mode: symptom logging and education.
Continuous improvement
Active learning loop: clinician edits and rejections feed back into preference-tuning datasets. Outcomes registry validates long-term accuracy.
Evaluation framework
| 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 expert clinician would suggest. |
| 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 and dosing must be deterministic. |
Safety and trust
Bounded scope
The system addresses menopause-related decisions only. Outside-scope questions are deflected with explicit handoff.
Human-in-the-loop by design
No autonomous prescribing. No autonomous patient messaging without clinician review. Recommendations only.
Explainability is a feature, not a setting
Every recommendation links to its evidence and inputs. If we can't show the work, we don't show the recommendation.
Bias monitoring
Sub-group performance tracking by age band, race, ethnicity, geography, and clinical setting. Reported quarterly to the clinical advisory board.
Red-teaming and adversarial evaluation
Quarterly red-team exercises by external clinical evaluators. Findings published internally and acted on before each major release.
Privacy posture
HIPAA-ready architecture. HITRUST CSF on the roadmap. PHI never leaves customer-controlled VPC in federated deployments.