Case Study: Ariv Health

A Healthcare Intelligence Platform, Built From Zero

Ariv Health sells market, referral, and financial intelligence to post-acute and value-based care organizations. PracticeVantage is the engineering team behind it: we designed the architecture, built every product, and operate the platform in production.

The Problem

Healthcare operators drown in vendor dashboards they cannot trust. Ariv's premise was different: every number traces back to verified CMS claims data, every product speaks the language of the people who run care organizations, and everything lives on one data foundation. Delivering that premise is an engineering problem before it is an analytics problem: Medicare claims at national scale, tenant isolation strict enough for competitors to share a platform, and payment mathematics that must match CMS methodology exactly.

What We Built

• A multi-tenant platform core: authentication, entitlements, data lake APIs, ingest pipelines, and metering, shared by every product.
• A product family on that core: home health, hospice, and skilled nursing intelligence, ACO strategy modeling with certified shared-savings math, specialty network benchmarking, a healthcare CRM, and a native iOS app.
• Claims-scale data engineering: CMS public and restricted files ingested, normalized, and served as precomputed read models for sub-second product experiences.
• An AI layer designed for regulated data: nine production MCP servers exposing governed analytics to AI assistants (including a remote server live in claude.ai), and an LLM service with hard per-tenant spend ceilings.

The Architecture

Serverless-first on AWS, Go end to end, everything managed as code. Two decisions define the design: reads are precomputed into per-tenant artifacts served from the edge, and every AI-facing path crosses a masking layer that strips PHI and suppresses small cells before any model, internal or external, sees a byte.

Engineering Decisions Worth Stealing

• Cost as a design property: AI features carry structural daily quotas per user and tenant, so the worst-case bill is known before launch, not discovered after.
• Correctness gates: actuarial formulas are frozen with subject-matter experts and enforced by byte-identical regression tests; a calculation cannot drift silently.
• Privacy as code: the PHI masking layer has its own regression pins, so a future refactor cannot quietly weaken it.
• Tenant isolation as a ship-blocker: row-level tenancy plus build-time leak gates that fail the release if tenant data is baked into any artifact.

In Production Today

Live products serving home health, hospice, ACO, and specialty network customers; the remote MCP server registered in claude.ai answering healthcare-analyst questions with policy and persona controls; and ingest pipelines that turn each CMS data release into refreshed intelligence without manual intervention.

Where The Architecture Goes Next

The roadmap extends the same foundations: more specialty networks on the benchmarking engine, deeper AI-analyst experiences over the governed data plane, and a wider MCP surface as enterprises adopt AI clients. Roadmap means roadmap: we distinguish what runs today from what comes next, here and with every client.

What This Proves

If your organization needs a data platform, an AI capability, or a regulated-industry system on Salesforce, Databricks, or AWS, this is the standard your project gets: architecture accountable to auditors, costs fixed by design, and software that runs for months without drama.