Your batches fail at 5–7%. We make that number zero.
AI-powered quality copilot that detects deviations before they become failures — and writes the investigation report for you.
Or try the pack verifier → — no signup, drop any BatchCortex pack ZIP and verify it in your browser.
🔒 EU-sovereign · GMP Annex 11 compliant · Data stays in Europe
Average cost per deviation investigation
Industry average (ISPE baseline)
Global pharma batch failure rate
BioPlan Associates, 434 facilities
Tablets produced between manual checks
Standard 30-min sampling at 100K tablets/hr
Watch BatchCortex work in real time.
No login required. Click the button to simulate a punch wear failure.
Batch BC-2041 | Paracetamol 500mg
Tablet Press Line 3 — 45 stations
Station Compression Forces (45 stations)
Live Sensor Readings
ML Ensemble Status
Press “Simulate Deviation” to watch the 11-layer ensemble detect a punch wear failure in real time
This is exactly what your QA team would see. Except it would be your batch, your equipment, your data.
Calculate Your Savings
Adjust the sliders to see how BatchCortex impacts your bottom line.
Estimated Gross Savings
€825K
Based on Analyze tier (€140/batch, €48,000/year annual minimum, €18,000 one-time setup) and a 30% reduction in batch failures through 15–30 minute early drift detection. Above the annual minimum, every additional batch is pure per-batch usage. Actual results vary by process type and facility.
The Hidden Tax on Every Batch
Quality failures are not rare events. They are a systemic cost baked into every production run.
1,000 deviations/year
Your QA team spends 6,000+ hours investigating. Most are documentation errors.
€50K–€5M per failed batch
Raw materials lost. Production time gone. Release delayed weeks.
4–8 hours per batch record review
Manual review is the bottleneck between production and release.
250,000 tablets between checks
Your operators sample 20 tablets every 30 minutes. That’s a 0.02% sampling rate. The other 99.98% are unchecked.
Current batch failure flow
How BatchCortex Works
Four steps from first deviation to QP signature — every batch, every station, every second.
STEP 01
DETECT
11-layer ML ensemble monitors every station, every tablet, every second. CUSUM catches 1σ shifts in 10 readings — while your Shewhart chart needs 44. That's 18 minutes of early warning your team doesn't have today.
STEP 02
ANALYZE
Contribution plots rank the top 5 features driving the alarm. Station 23 punch wear vs feeder blockage — diagnosed in under a second, not hours. Every alert arrives with full sensor, model, and operator context attached.
STEP 03
ESCALATE
SMS to your QA Manager in under 3 seconds. No response in 3 minutes? QA Director gets the same alert. Still nothing? An automated phone call reads the batch status aloud. Four tiers, zero latency.
STEP 04
DOCUMENT
Auto-generates ALCOA+ deviation reports, batch-record annotations, and CAPA drafts in under 60 seconds — ready for QP signature. Chained SHA-256 signatures plus point-witnessed audit events. Inspector-ready on every export.
Equipment Integration
Connects to your existing systems. No modifications required.
BatchCortex connects to your existing manufacturing systems via industry-standard protocols. Your equipment does not need to be modified. We read data from your control systems in read-only mode — we cannot write to your production equipment.
OPC-UA
Industry standard. Covers 60–70% of pharma DCS/SCADA systems (Siemens, Rockwell, ABB, Emerson DeltaV). Auto-reconnects with exponential backoff if the connection drops.
MQTT
IIoT setups, Siemens MindSphere, and modern sensor networks. Full anomaly detection pipeline — same AI as OPC-UA mode.
The richest per-batch evidence record in pharma AI.
SIMCA gives you PCA residuals. Aizon gives you dashboards. BatchCortex gives the inspector the raw machine truth — eight parallel evidence streams captured at the source, timestamped at the OPC-UA server, chained into an append-only audit trail, and archived into every Inspection Readiness Pack.
PackML state trio
Every PackML state transition subscribed at the source — CurrentState, UnitMode, UnitModeChangeInProgress — chained into the batch audit trail. Inspectors can reconstruct the exact machine state at any second of the run.
HMI audit witnesses
AuditSessionEventType, AuditWriteUpdateEventType, AlarmConditionType, and ProgramTransitionEventType subscribed from the control system. Every operator login, parameter edit, alarm acknowledgement, and program switch is captured with its ClientUserId.
Per-outcome model confidence
Every anomaly score carries its confidence value, decision threshold, and the exact model_hash that produced it. Annex 22 draft §5.3 compliance baked in — no averaged confidences, no model ambiguity.
Operator override delta
When an operator disagrees with the model, the recommended action and the actual action are both stored — with attribution, timestamp, and rationale. The full model-vs-human delta is an Annex 22 reviewable record.
Tablet rejection stream
Every reject-chute event from PackML presses logged with root-cause code, station, recipe step, and source timestamp. No more sampled quality data — every single rejected tablet is in the batch record.
Recipe + Established Condition
Recipe id, version, and the ICH Q12 Established Condition flag captured on every sensor reading. Post-approval change inspectors see exactly which parameters were regulatory-committed and which were not — at the row level.
Edge session + machine identity
Every reading tied to an edge_session_id plus the machine's OPC-UA identification block (vendor, model, serial, firmware, hardware revision). If the same machine ran under two edge agents, the seam is explicit in the data.
Scheduled audit-trail reviews
Annex 11 §9 mandates periodic audit-trail review. BatchCortex scheduler creates reviewer-signed review events at the configured cadence, each one a chained witness in events_log — no more spreadsheet logs of who-reviewed-what-when.
Every stream above is ALCOA+ Original, server-side timestamped, quality-coded, and archived into the Inspection Readiness Pack — no sampling, no aggregation, no summary-only rows.
Detect. Explain. Escalate. Prove it.
No single model is trustworthy enough for a GMP batch. BatchCortex combines a ten-layer detection ensemble with SHAP-level explainability, a four-tier escalation ladder, and a cryptographically chained audit trail — so every alert is defensible in front of a regulator, not just a dashboard.
Detection · 10 parallel layers
Ten algorithms. One verdict.
SIMCA-class tools give you one number from one PCA model. BatchCortex runs ten detection layers in parallel — each targeting a different class of failure — and fuses them into a weighted ensemble score you can trust.
10×
Independent models
CUSUM
Cumulative sum control — catches small, persistent drifts long before any single reading crosses a threshold.
EWMA
Exponentially weighted moving average — tracks slow trajectory shifts in dwell time, force, and fill depth.
Hotelling T²
Distance from the PCA-defined normal operating region. Spots when the joint process state drifts, even if each sensor looks fine.
SPE (Q-statistic)
Squared prediction error outside the PCA model — flags failure modes the model has never seen before.
Isolation Forest
Unsupervised outlier detection across all feature dimensions. No labels needed — trained directly on your good batches.
TCN Autoencoder
Temporal convolutional network reconstructs normal sensor sequences. High reconstruction error = anomaly in the time domain.
LSTM Autoencoder
Recurrent sequence model for long-range dependencies. Catches anomalies a purely feed-forward model would miss.
Wavelet Energy
Frequency-domain analysis on high-rate compression signals. Detects mechanical vibration shifts invisible in the time domain.
Cross-Correlation
Tracks the coupling between sensors. When force and dwell time stop moving together, something upstream has changed.
Changepoint
Bayesian changepoint detection — locates the exact moment the process entered a new regime, not just that it did.
Plus a parallel survival model producing time-to-failure estimates for every running batch — so your QA team knows how many minutes they have, not just that something is wrong.
Explain
SHAP attribution, per alarm
Every anomaly comes with a plain-English reason: which sensor moved, in which direction, and how much it contributed to the alarm. No black boxes — Annex 22 and EU AI Act explainability built in.
Escalate
Four-tier escalation ladder
Quiet warning to emergency call, with HMAC-signed one-click acknowledge links at every step. No alert slips through the cracks — and every escalation is logged to the audit trail.
Shift QA operator
QA manager
Qualified Person (QP)
All QA + QP + Admin
Prove
Chained-hash audit trail
Every sign-off is SHA-256 hashed and chained to the previous audit entry. The events log is append-only at the database level — enforced by Postgres, not application code. Tamper one record and the chain breaks.
Everything Your QA Team Needs
Detection is just the start. BatchCortex handles investigation, documentation, quality reviews, and model governance — so your team focuses on decisions, not paperwork.
AI Advisory
Ghost Operator
Real-time severity classification, root cause hypothesis, failure mode prediction, and recommended actions — all with confidence scores and evidence. Your AI co-pilot for every anomaly.
Annual PQR
Product Quality Review
Cpk process capability, Nelson rules, OOS/OOT detection, linear regression, and deviation severity — all computed automatically. AI writes the executive summary. Period-over-period benchmarking included.
Audit-Ready Archives
Inspection Readiness Pack
One click builds a cryptographically sealed ZIP of every batch's evidence — cover PDF, per-batch records, raw sensor CSVs, events log, and signatures. Merkle-rooted, hash-chained, and verifiable offline through an embedded verify.html. From four-week audit scramble to ninety seconds.
Side-by-Side Analysis
Batch Comparison
Compare any two batches with station-level heatmaps, ensemble score deltas, and anomaly event timelines. Instantly spot what changed between a golden batch and a failing one.
Context-Aware Q&A
AI Investigation Chat
Ask questions about a live or completed batch and get pharma-specific answers grounded in real sensor data, ML scores, and event history. Like having a process engineer on call 24/7.
Lifecycle & Drift
Model Registry
Full version history of every deployed model — training metrics, SHAP charts, deployment timestamps, and change control references. PSI drift monitoring triggers retrain recommendations automatically.
Continuous Improvement
Training Feedback Loop
QA operators label anomalies as true or false positives. Dismissal patterns are detected automatically and surfaced to admins. Every label makes the next model smarter.
Your batch data stays in Europe.
While most AI platforms route your sensitive manufacturing data through US servers, BatchCortex runs on Swedish infrastructure. Your process data, your IP, your patients — all protected under EU jurisdiction.
Hosted in Sweden
Production AI runs on Mistral AI — EU-incorporated, based in France. EU data residency guaranteed. GDPR compliant by architecture, not by policy.
Sovereign from day one
BatchCortex was designed for the EU AI Act before most vendors knew it existed. Annex 22 compliant. No US data transfers for production workloads.
No vendor lock-in
Built on open-weight European AI models. Mistral Large via Swedish data centers from 2027. Your compliance doesn't depend on a San Francisco boardroom.
Developed in Sweden · The Silicon Valhalla · EU AI Act compliant · batchcortex.com
Built for GMP from Line One
Regulatory-ready. Not an afterthought.
Every feature is designed with pharmaceutical compliance requirements at its core.
EU GMP Annex 11
Compliant
EU GMP Annex 22 (AI)
Ready
21 CFR Part 11
Designed for
GAMP 5 Category 5
Validation package in development
ALCOA+ Data Integrity
Compliant
EU AI Act
Limited risk — classified
Sartorius SIMCA-equivalent
PCA/T²/SPE monitoring
Human-on-the-loop architecture. Every AI recommendation requires operator approval. Full audit trail. No black boxes.
Same math. Deeper intelligence.
We use the same PCA/T²/SPE monitoring that SIMCA pioneered — plus five additional detection layers that catch what multivariate statistics alone cannot.
Traditional MVDA (SIMCA, etc.)
BatchCortex
PCA on batch-level averages
11-layer ensemble on per-station, per-second data
Tells you a batch went wrong
Tells you WHICH station, WHICH failure mode, and HOW LONG until spec breach
∼€50K license + training
Pay-per-batch, no license
General-purpose toolkit
Purpose-built for tablet compression
Manual model building
Auto-calibrating ensemble
No temporal awareness
TCN + LSTM autoencoders detect drift patterns
SIMCA is the gold standard for multivariate batch analysis. BatchCortex builds on that foundation with purpose-built layers for tablet compression — same statistical rigor, deeper real-time intelligence.
Pay Per Batch. Not Per Promise.
Per-batch pricing with a predictable annual floor. Your production volume drives your spend — and your CFO can forecast it.
Monitor
min €24,000/year
+ €12,000 one-time setup
Real-time visibility into your batch process.
- OPC-UA & MQTT connectivity
- Real-time parameter monitoring
- Hard limit & drift detection
- Automated SMS/email escalation
- Immutable audit trail
- Annex 11 compliant dashboard
Analyze
min €48,000/year
+ €18,000 one-time setup
AI-generated GMP deviation reports and electronic sign-off.
- Everything in Monitor
- AI deviation report generation (<60s)
- Electronic sign-off (21 CFR Part 11)
- Ghost Operator AI advisory
- Annual PQR generation
- Inspection Readiness Pack (unlimited)
- Scheduled monthly/quarterly pack generation
- Batch comparison & heatmaps
- AI investigation chat
- QA feedback loop & model improvement
- LIMS at-line QC integration (CSV + adapter)
- Material genealogy & recall trace (API lot to batch)
- IQ/OQ validation document support
- EU-sovereign AI inference
Guard
min €78,000/year
+ €24,000 one-time setup
Full GMP compliance stack with dedicated onboarding.
- Everything in Analyze
- NIR PAT blend uniformity ingestion
- Annex 22 NIR input-drift monitoring
- Model registry & drift monitoring
- Training data pipeline & snapshots
- Custom OPC-UA/MQTT integration support
- Dedicated onboarding engineer on-site
- Custom alert thresholds & escalation rules
- Priority support (4-hour response)
- Multi-line monitoring (unlimited lines)
Mandatory one-time setup per site. No multi-year lock-in. Pilot partners get 50% off per-batch rate, annual minimum, and setup fee.
Frequently Asked Questions
Common questions about BatchCortex, GMP compliance, and our pilot program.
First 3 pilot partners get 50% off setup — and a custom ML model trained on your press data
50% off the one-time setup fee in exchange for a brief written case study after 90 days. EU AI Act enforcement begins August 2026 — validate your AI pipeline now and gain a 12-month compliance head start. Secure your pilot spot before ISPE Copenhagen.
Vilmer Frost
Founder, BatchCortex — Stockholm, Sweden
Built by a developer who spent months studying EU GMP regulations before writing a single line of product code. BatchCortex exists because the compliance gap is real and the existing solutions are too slow, too expensive, and not built for EU data sovereignty.
Get in Touch
Have a question about BatchCortex, Annex 22 compliance, or pilot partnerships? We'd love to hear from you.