AI Development Company in Dublin

Pharmaceutical Manufacturing

A global pharmaceutical company's Irish manufacturing site (producing biologics — monoclonal antibodies for oncology and autoimmune diseases — 1,200 employees, €2.8 billion annual production value) was losing €45 million annually to batch failures and quality deviations. The biologic manufacturing process involves cell culture in 15,000-litre bioreactors — a 14-day process where living cells produce the therapeutic protein. Despite extensive in-process controls, 8% of batches failed final quality testing (out of specification for critical quality attributes — potency, purity, or aggregation levels). Each failed batch cost approximately €3.2 million (raw materials, utilities, labour, and 14 days of bioreactor capacity). The site had attempted simple statistical process control (SPC) — monitoring individual parameters against specification limits — but SPC only caught deviations after they occurred, couldn't predict batch outcomes from early-stage measurements, and couldn't identify the complex multivariate relationships between process parameters and quality outcomes.

We built a batch quality prediction system that forecasts final quality attributes from in-process measurements, enabling early intervention. (1) Process data integration: connected to the site's process historian (OSIsoft PI) and MES (manufacturing execution system). Data sources: bioreactor sensors (temperature, pH, dissolved oxygen, agitation speed, glucose feed rate, lactate levels — measured every 30 seconds throughout the 14-day culture), offline analytics (cell density, viability, titre measurements — taken daily by lab analysts), raw material characteristics (cell bank passage number, media lot properties, buffer specifications — each batch uses specific raw materials whose properties affect outcomes), environmental data (cleanroom temperature, humidity, differential pressure — environmental conditions during inoculation and harvest affect quality), and historical batch data (1,200 completed batches over 4 years — including 96 failed batches that provide critical training signal for predicting failure modes). Total: 340 process parameters per batch, measured at various frequencies over 14 days, plus 45 raw material characteristics and 12 environmental factors. (2) Time-series batch modelling: traditional ML treats each batch as a single data point. We implemented a time-series approach that uses the temporal evolution of process parameters — not just point values but trajectories. The model learns: normal process trajectories (how temperature, pH, and dissolved oxygen should evolve over the 14-day culture — healthy batches follow characteristic profiles), early deviation signatures (subtle deviations in days 1-5 that predict failures on day 14 — the model identifies early warning patterns invisible to human operators monitoring individual parameters), multivariate relationships (interactions between parameters — a combination of slightly high pH and slightly low dissolved oxygen during the exponential growth phase that individually fall within specifications but together predict quality risk), and raw material effects (certain combinations of cell bank passage number and media lot properties create quality risk — enabling proactive raw material screening). Model architecture: LSTM (Long Short-Term Memory) network for the time-series evolution, combined with a gradient-boosted tree for static batch features (raw materials, environmental). The combined model predicts: probability of batch success (continuous 0-1 score updated daily), predicted final potency (with confidence interval), predicted purity profile, and risk decomposition (which parameters are contributing most to any identified risk — enabling targeted intervention). (3) Early warning system: the model generates daily predictions from day 3 onwards (the earliest point with sufficient data for meaningful prediction). Day 3 prediction accuracy: 72% (binary success/fail). Day 7 prediction accuracy: 88%. Day 10 prediction accuracy: 94%. When the model predicts risk, it generates: risk level (low/medium/high/critical), contributing factors (ranked list of parameters driving the risk prediction — "dissolved oxygen trajectory deviation is the primary risk factor, contributing 34% of predicted risk"), recommended actions (process adjustments that could mitigate the risk — "increase dissolved oxygen setpoint from 40% to 50% for the next 48 hours"), and historical analogs (similar past batches — showing which ones recovered with intervention and which ones failed — giving operators empirical evidence for the recommended action). (4) GAMP 5 validation: the AI system operates in a GMP environment and was validated according to GAMP 5 category 5 (custom software). Validation deliverables: User Requirements Specification (URS), Functional Specification (FS), Design Specification (DS), Installation Qualification (IQ), Operational Qualification (OQ), Performance Qualification (PQ), and traceability matrix linking requirements through testing. All predictions are logged immutably with timestamps, model version, and input data — creating a complete audit trail for HPRA and FDA inspections. (5) Human-in-the-loop: the AI recommends — humans decide. Process engineers review model predictions daily. No automated process changes — all adjustments require engineer approval with documented rationale. The system includes "model confidence" indicators that clearly communicate when predictions are high-confidence (based on well-represented historical patterns) vs. low-confidence (unusual parameter combinations not well-represented in training data).

Batch failure rate: decreased from 8% to 2.8% (65% reduction). The model identified at-risk batches early enough for process engineers to intervene in 62% of cases that would have otherwise failed. Annual savings: €45M × 65% = €29.3M in prevented batch failures. Each prevented batch saves €3.2M in direct costs — 9.2 batches saved annually. Early detection: the day-7 prediction (88% accuracy) is the primary decision point. Process engineers now review model predictions at the day-7 milestone for every batch — a practice that has become standard operating procedure. 78% of successful interventions occur between days 7-10. Prediction accuracy: day-3: 72%, day-7: 88%, day-10: 94%, day-12: 97%. The model improves as the batch progresses and more data becomes available. For high-risk predictions on day 10, the model has 94% accuracy — a sufficiently high confidence level for engineers to make significant process adjustments. Root cause insight: the model identified two previously unknown risk factors. A specific media lot supplier produced raw materials that created a 3x higher batch failure rate when combined with late-passage cell banks — this interaction was invisible in single-parameter SPC but clear in the multivariate model. The site changed supplier qualification procedures based on this insight, preventing an estimated 4 additional failures annually (€12.8M). HPRA inspection: the system passed HPRA GMP inspection with no findings. The GAMP 5 validation documentation and audit trail were specifically commended. The inspector noted that "the quality prediction system demonstrates an advanced approach to process understanding consistent with ICH Q8 principles." Scale-up: following the success at this site, the pharmaceutical company is deploying the system to three additional manufacturing sites (Singapore, Puerto Rico, and County Cork) — total programme value estimated at €80M annually across four sites. Development investment: €480,000 over 20 weeks.