Architecture

System of insight, not system of record.

Your historian stores what happened. Praescia tells you what it means. Five layers. Three Atlas engines. One GxP compliance shell. Every architectural choice documented, measurable, and reproducible.

The fundamental distinction

OSIsoft PI, Honeywell PHD, Rockwell FactoryTalk, Werum PAS-X, LabVantage — every manufacturing IT system at a biopharma site today is a system of record. It stores batch data, logs events, tracks samples, automates the process record. Each is a silo. Each sees only what it was built to see.

Praescia is a system of insight. It reads from every system of record the site already runs — without replacing any of them — fuses signals into a live batch state, detects the onset of deviation or constraint, advises the operator, and measures its own accuracy against real batch outcomes.

The shift changes the operating posture of the manufacturing site. Instead of process engineers manually correlating historian trends across multiple screens, the intelligence layer does it continuously. Instead of OOS risk depending on the release assay, it is projected hours before. Instead of a deviation report being written after the loss, the transition is caught on onset — while there is still time to act.

The five layers

L5 · Verticals

Upstream bioprocess · Fill-finish · Cell & gene therapy · CDMO fleet

Each vertical is a domain vocabulary + golden-trajectory profile + monitor + skin. New verticals do not require core-code changes. The same spine underlies every one.

L4 · Read-only source adapters

Ingest · Normalize · Validate · Publish

Every integration is a plugin, not a fork. Standards-based (OPC-UA, MQTT, REST, SFTP) + biopharma-vendor connectors (OSIsoft PI, Rockwell, Werum PAS-X, LabVantage, Benchling, Veeva). Data is read, never modified, never replicated into an external store.

L3 · Combining Layer + Knowledge Graph

16 node types · 14 edge types · the genuinely new IP

Batch · BioreactorRun · ProcessPhase · Signal · GoldenTrajectory · Deviation · Advisory · Disposition · Equipment · Line · Utility · CleanroomEvent · ColdChainLeg · TechTransferEvent · ValidationRecord · AuditEntry. Identity resolution, signal fusion, state fusion, cross-site benchmarking, contradiction detection, unified recommendation. Produces conclusions no single tool could reach.

L2 · Biopharma State Spine

Seven-stage deterministic pipeline

OBSERVE → NORMALIZE → DETECT_STATE → DETECT_TRANSITION → ESTIMATE_OUTCOME → RECOMMEND → VALIDATE. Composed from proven Atlas PC (production constraint) and Atlas OI (operations intelligence) engines. Rule-based classifiers, stdlib-only Python, zero third-party runtime dependencies.

L1 · Manufacturing data

Read-only adapters. One-way ingress.

Process historians · In-line sensors · MES batch records · LIMS sample data · Environmental monitoring · Utility systems (WFI, HVAC) · ELN entries · Batch disposition records · Tech transfer packages · Cold-chain telemetry. Nothing modified. Nothing written back. The data stays yours.

Engine composition (bill-of-materials)

Praescia is a tier-3 Atlas superstructure — it composes proven Atlas engines rather than building from scratch. The same coupler and “Atlas Inside” network layer that powers the sibling custra (smart buildings) and Connexum verticals, re-skinned for biopharma.

Role	Engine	What it contributes
Core — batch intelligence	Atlas PC (Production Constraint Spine)	Bioreactor / fill-finish constraint state, golden-batch deviation detection, bottleneck-transition, throughput advisory
Core — operations intelligence	Atlas OI (Industrial Operations)	Plant / utility / cleanroom operating state (WFI, HVAC, EM, equipment health), operational-transition + intervention advisory
Extension — cold-chain	Atlas MO (Movement Optimization)	Material and sample logistics, cold-chain excursion risk, warehouse coordination
Coupler	`atlas_compose`	Couples engines into one pipeline (Atlas root); same coupler as every other Atlas vertical
Network layer	“Atlas Inside”	Cross-site / cross-line density layer; fleet benchmarking; the breadth monetization tier
Praescia delta	GxP compliance shell	21 CFR Part 11 / EU Annex 11 / GAMP 5 / ALCOA+ wrapper: audit-trail, e-signatures, data-integrity, validation package

Core = Atlas PC + Atlas OI. Atlas MO is additive per deployment. The same coupler and network layer across every Atlas vertical — the reusable machinery is unchanged.

The domain vocabulary — what the spine detects

Upstream bioprocess (Atlas PC core)

Bioreactor batch intelligence

Observe/Normalize: pH, DO, temperature, agitation, biomass/OD, feed rate, off-gas, metabolites
State: lag / exponential / stationary / decline · on-golden-trajectory vs. drifting
Transition: onset of deviation (metabolic shift, DO crash, foaming, feed deviation)
Outcome: predicted titer / yield / batch-success probability, called pre-harvest
Advisory: feed / setpoint / harvest-timing recommendation (insight-only)

Fill-finish / drug product (Atlas PC + OI)

Line constraint intelligence

Observe/Normalize: line telemetry, reject counters, EM data, utility state
State: constraint / bottleneck state of the line; cleanroom operating state
Transition: bottleneck onset, EM excursion onset, changeover risk
Outcome: throughput / yield-loss / batch-record-risk estimate
Advisory: explainable intervention advisory to the operator

Deployment postures

Cloud tenant

Managed SaaS in customer-selected cloud region (AWS, Azure, GCP). Customer-controlled keys via cloud KMS. Ideal for CDMOs and biotech without on-premises infrastructure requirements.

On-premises

Containerized appliance in the customer’s own data center or DMZ. Air-gappable. Data never leaves the plant network. Required for the most regulated GMP environments.

OEM / embedded

Intelligence licensed into a bioreactor, PAT instrument, or MES vendor’s product stack. The “Atlas Inside” mark. Available to bioreactor and PAT hardware partners.

Why composition, not a build from scratch

Every stage of the Biopharma State Spine already exists as a shipped Atlas capability. Biopharma manufacturing is the same shape as the industrial domains Atlas already serves: telemetry → state → transition → outcome → advisory. Praescia is the re-skin + the GxP compliance shell — the only thing the sibling Atlas verticals don’t carry.

Lower build risk. The engines are proven in adjacent manufacturing verticals. The composed pipeline (Atlas PC + Atlas OI via atlas_compose) passed 73 tests on a synthetic bioprocess rig before a single real batch arrived. The evidence gate is what we’re working on next — not the plumbing.