The AI operating system for pharmaceutical manufacturing.
Verify what your contract manufacturers send, investigate deviations with the evidence in view, and forecast process outcomes with physics-based models. Every proposed change waits for a named person to sign.
Working now: CMO delivery verification on real CHO data, the assistant and freeze-drying forecasts. Monitoring and investigation screens preview a synthetic campaign.
Manufacturing teams
use it for
- CMO deliveries
- Batch monitoring
- Deviation investigations
- Root-cause evidence
- Physics-based forecasts
- Freeze-drying design
- Yield optimization
- Tech transfer
- Quality reviews
- Lab results
- Knowledge search
- Operator shifts
- Tech transfer data
The problem
Manufacturing data lives in a dozen systems. Engineers still connect it by hand.
What that leaves you with —
The answer is spread across systems
Batch records, historian tags, lab results and work orders each hold part of the story. Every investigation starts with exports and spreadsheets.
Investigations start from scratch
Each deviation means rebuilding the same timeline by hand, and the reasoning ends up in slide decks that are hard to trace back to the data.
Changes get tested on real batches
Without a model that follows the process physics, a new site, scale or recipe is checked on the plant itself, one batch at a time.
The platform
One operating system from the plant floor to the signature. Every claim one click from its source.
Connect
Verified CMO deliveries today; read-only access to MES, historian, LIMS and QMS records. Every value keeps its source and timestamp.
Investigate
Agents assemble timelines and candidate causes, with the evidence for and against each one labelled.
Predict
Physics-based process models forecast what a change will do, with their inputs and assumptions on record.
Act
Findings become proposals with an owner, expected effect and required signatures. Nothing executes on its own.
From partner data to signed decision.
October delivery · 2,555 source elements
real CHO dataMeasurements1,843
Identifiers596
Flagged for review116
Unaccounted for0
pH 7.033 · CHO-004-R01 · 0 hxlsx · Process data!C2
Physics-based models
Predictions that follow the physics, tested in the open.
Statistical models learn from the batches you have already run. Physics-based models also obey mass and heat balances, so they can forecast conditions you have not tried: a new site, a new scale, a new recipe. Each forecast records its model version, inputs and assumptions, and the agents use these models as tools.
Freeze drying
Primary drying: heat transfer to the vial, mass transfer through the dried layer, ice vapor pressure and condenser capacity.
Reconstructed demonstration · not qualified
Penicillin fed-batch
Industrial-scale growth, production, oxygen transfer and substrate balances (IndPenSim / PenSimPy).
Scored on 96 h · full 230 h batch not qualified
Antibody upstream
A 17-state reactor and separator model of a continuous monoclonal antibody process (SMPL).
Simulation benchmark
CHO batch culture
Glucose-limited growth and death with a substrate balance, fitted to real CHO-K1 measurements.
Tested on real data
The same AI, with and without a physics model.
48 simulated situations per process, two independent sessions per arm, 1,136 of 1,152 attempts scored. Both arms saw identical data and limits; only one could run the model.
Antibody upstream · recipe adaptation
+12.9 points, 95% interval 4.5 to 23.1
82.5→95.5
Penicillin process · recipe adaptation
+10.3 points, 95% interval 4.0 to 18.0
86.3→96.6
Penicillin process · operating decisions
+8.2 points, 95% interval 3.0 to 15.0
88.4→96.6
Antibody upstream · operating decisions
+2.5 points, 95% interval −1.8 to 8.0
94.2→96.7
Forecasts
Mixed. Forecast error fell 9.6% for the antibody process and rose 11% for penicillin. Neither met our pre-set bar of a 20% reduction.
Real CHO cultures
The simple growth model made glucose forecasts less accurate (mean error 0.60 → 0.72 g/L). A model this simple leaves out nutrients and stress the cells respond to.
What this shows
In simulation the tool and the scorer share equations, so gains show the value of a correct model, not plant accuracy. No interval cleared our pre-set threshold. Model corrections and a fresh run are under way.
Working now · CMO deliveries
Verified data from your contract manufacturers, inside the same platform.
Drop in a delivery as it arrived. The platform reads each file, traces every value to its source cell and lists what it cannot account for. Once a delivery is verified, send it to the assistant: it answers with the batch, time and source file behind each value.
Measured on 24 real CHO bioreactor runs from a published dataset, repackaged as monthly deliveries from a fictional CMO.
100%
Values verified automatically on a repeat delivery
1,843 of 1,843
0
Wrong values accepted without review
scored against independent answer keys
99.7%
Verified automatically after a layout change
after one confirmation
0
Source cells left unaccounted for
every condition tested
Deterministic local processing, no model calls. Defects in the changed delivery are injected and listed. Human time saved has not been measured.
How it works
Starts from the systems you have, without replacing any of them.
No migration and no rip-and-replace. The platform reads from the systems you have already validated, and they stay the record.
01
Start with the data you receive
Drop in a CMO delivery today, or connect your MES, historian, LIMS and QMS read-only. Nothing is written back, and every value keeps its source, version and timestamp.
02
Start from a question or an event
A deviation, a yield drop, a transfer or a partner delivery. Agents gather the evidence and show each step they took.
03
Review, sign, then act
Proposals carry a rationale, expected effect and required approvals. Implementation stays in your systems, with an audit trail designed to support 21 CFR Part 11.
Validation of any deployment remains the customer’s responsibility. The platform does not claim GxP validation.
Systems
Built around the systems you already run.
Manufacturing and quality systems
Read-only connections that index and link records without changing them. Each source stays authoritative.
DeltaV eBR
Batch records
OSIsoft PI
Historian
LabWare LIMS
Lab results
Veeva Vault QMS
Deviations, CAPA
Maximo CMMS
Work orders
Documentum
SOPs
The public demo shows these connections with synthetic records.
Process models
Registered physics-based models run beside your data. Inputs, versions and assumptions are recorded with every forecast.
CMO deliveries · working now
Excel exports, LIMS CSVs, PDF certificates and ZIP packages are read, checked and traced to their source files, then shared with the assistant as evidence.