The Factory — Overview¶
Platform → Engine → Agent → Disease Program. The HCLS AI Factory is one open-source system in four layers: horizontal compute engines, clinical reasoning agents over them, disease programs that compose them for one condition, and the platform glue that governs it all.
A three-minute guided tour of the whole factory — narrated and captioned. Decision support, not diagnosis.

Illustrative architecture diagram.
The model¶
- 8 Engines — the horizontal compute muscle: genomics, clinical interpretation, therapeutic discovery, imaging, oncology, cardiology, large-molecule / structural biology, and single-cell.
- 8 Intelligence Agents — the clinical reasoning layers that interpret what the engines produce. Each is decision support for a qualified clinician — never autonomous diagnosis or prescribing.
- 1 Disease Program — Tuberous Sclerosis Complex, the flagship vertical that composes the whole factory for one child in one governed afternoon.
- One Platform — the Capability Registry (the typed source of truth), the MCP tool-surface, the Workflow Composer, single-box MLOps, and the governance gates.
- Frontier Models — heavy or partner-gated models the factory connects to on demand via elastic burst (Anthropic's Claude, NVIDIA BioNeMo, and Chai Discovery, plus the roadmap frontier), each with an honest status.
Engine vs. agent — the rule of thumb
An engine turns inputs into a result — a variant file, a docked molecule, a segmented scan. An agent reasons over those results the way a specialist would — a pharmacologist checking a dose, a geneticist ranking diagnoses — and always hands the decision back to the clinician. Engines are the lab; agents are the consult. And "governed" means every step runs behind the platform's governance gates — logged, access-controlled, and honesty-checked — so that afternoon is auditable, not a black box.
See it honestly¶
- The Capability Maturity Matrix shows every capability's real, registry-declared status — the site cannot claim ahead of the code.
- The Capability Brief is the full picture, in a technical cut and a mission cut.
How to read this
Decision support, not diagnosis. "Elastic burst," not "all on one box" — heavy or
ARM-incompatible models run on remote GPUs over a private mesh. Real, never mocked — a
capability marked live is served by a real model against real input.