How a small team used AI throughout the build to surface over $10M in operational savings on the first pass.
Hospital organizations sit on enormous amounts of operational data — process logs, supply chain records, scheduling, inventory, billing, contract terms — but the data is scattered across systems that don't talk to each other and structured (when it's structured at all) for billing, not for decisions.
ArcheHealth's mission was to help these organizations find cost savings and new revenue opportunities buried in that data — not by adding another dashboard, but by surfacing specific, actionable insights an executive could act on next week.
Doing that meant building a real platform: ingestion across heterogeneous systems, statistical analysis layered with AI reasoning for the messy parts, and a delivery model that produced findings hospital leadership trusted enough to act on. With a small team. Quickly.
The interesting move wasn't using AI in the product — that was table stakes. The interesting move was using AI through the build itself, which is how a small team shipped a real platform fast enough to matter. This engagement was the prototype of what's now productized as AI-Native Development — methodology + operators + AI coding agents, replacing what would otherwise need a much larger engineering team.
Used Claude Cowork to translate stakeholder conversations into structured technical designs. Reduced the requirements-to-spec cycle from weeks to days. Stakeholders saw their words turned into diagrams in the same meeting.
AI-assisted across code generation, test generation, and infrastructure-as-code using Claude Code for the implementation work and Claude Chat for ad-hoc reasoning, debugging, and design decisions. Engineers stayed in their reasoning loop while the typing-and-boilerplate part collapsed. Code review remained human; generation didn't.
Hospital data lives in dozens of source systems. Databricks was the lakehouse — ingestion, transformation, feature engineering, and notebook-driven exploration all in one place. Statistical models and AI reasoning ran against curated tables with lineage and audit trails intact.
The product combined classical statistical models for what they're good at (regression, anomaly detection, baseline comparisons) with AI reasoning for the parts where context, framing, and "what's actually going on here" matter more than math.
Multi-LLM by design — different problems route to different models based on cost, latency, and what the eval harness says actually performs.
Across processes, supply chain, and operations, the first analysis run surfaced over $10 million in identifiable cost savings and revenue opportunities — specific enough that an operations executive could decide which to action and in what order.
More importantly: the platform was set up to do this continuously. Each new dataset, each new contract, each new operational change feeds back into the analysis. The first $10M was the proof. The next $10M is what the system is built to find on its own.
Two things stand out from this engagement that I'd take into the next one:
If your organization is sitting on operational data and AI feels like the right lens — let's talk about whether it actually is, and what shipping it would look like.