Won
Created 7/15/2026, 6:58:49 AM
Complexity
MEDIUM
Problems
2
identified
Data sources
1
classified
Agent runs
8
on this case
Discovery Analyst
Structured analysis
Hilltop Health operates a small clinic network and wants a RAG-based assistant for clinicians to query internal protocols. Single data source (Notion), strict HIPAA compliance, US-only data residency.
Objectives
- Reduce clinician time on protocol lookups from 4-8h/day to <30min/day
- Ensure HIPAA compliance end-to-end (PHI redaction + BAA + audit log)
- Provide 24/7 coverage (office manager not on shift nights/weekends)
Problems to solve (2)
PTS-1 · Clinician
Office manager is the bottleneck for protocol lookups
5-10 min per query, blocks patient care
PTS-2 · Compliance officer
No audit trail of protocol queries
Cannot demonstrate due diligence in case of incident
Open questions (1)
HIGH How much PHI actually exists in Notion pages vs attached clinical notes?
Solution Architect
Target architecture
Components
C-1 Clinician chat UI frontend
C-2 RAG service sveltekit-server
C-3 Notion sync worker sveltekit-server
C-4 Postgres + pgvector (US) infrastructure
C-5 Audit log sink infrastructure
Workflows (to-be)
WF-1 custom-code
Single-purpose RAG with strict compliance — easier to audit than n8n. Engineering team owns and can iterate fast.
Key decisions
Managed Postgres in US (Supabase US region)
HIPAA BAA available, pgvector included, US-only data residency
→ database-rules.md#rule-3
Custom RAG service (no n8n)
Easier to audit + BAA coverage + finer control over retrieval
→ n8n-rules.md#rule-3
OpenAI gpt-4o-mini with BAA + zero-retention
BAA signed, cost-effective for grounded RAG, low latency
→ security-patterns.md#hipaa
Vendor lock-in: LOW
OpenAI with BAA — swappable to Anthropic. Notion data is standard markdown. Standard SQL.
Cost Estimator
Investment
Person-days
24
18–30 days
Monthly
$320
operational
Support
$700
per month
Agent runs
Every agent invocation on this case, in order.
- Donediscovery-analystmock/mock-1 · 920→740 tok · 144ms
- Doneworkflow-analystmock/mock-1 · 1080→880 tok · 168ms
- Donesecurity-analystmock/mock-1 · 1020→810 tok · 192ms
- Donesolution-architectmock/mock-1 · 1280→1100 tok · 208ms
- Donecost-estimatormock/mock-1 · 820→640 tok · 126ms
- Doneproposal-writermock/mock-1 · 980→1380 tok · 220ms
- Donereport-generatormock/mock-1 · 740→1860 tok · 288ms
- Failedsecurity-analystopenai/gpt-4o · 1180→0 tok · 31200msRate limit (429) — openai fallback to anthropic succeeded