Self-Hosted AI Deployment in Barcelona, Spain

Pharmaceutical / Manufacturing / Self-Hosted AI / Barcelona

A Barcelona-headquartered pharmaceutical company (EUR 2.4B revenue, specialising in plasma-derived therapies, hospital injectables, and dermatology products, operating 12 manufacturing plants across Spain, Portugal, France, and Germany, 8,400 employees, 180+ products in 90 countries) needed to deploy AI capabilities to transform manufacturing quality, R&D productivity, and pharmacovigilance — while maintaining absolute control over data that included patient clinical information, proprietary manufacturing processes, and pre-patent drug development research. Core challenges: (1) Pharmaceutical data sovereignty — the company processed four categories of extraordinarily sensitive data. Clinical trial data: data from 28 active clinical trials involving 14,000 patients across 120 trial sites in 18 countries. Clinical trial data included: patient demographics, medical histories, adverse events, laboratory results, and efficacy endpoints. This data was governed by the Clinical Trials Regulation (EU 536/2014), ICH GCP guidelines, and each country's national data protection law. Any breach of clinical trial data could terminate trials, trigger regulatory action from EMA and national authorities (AEMPS in Spain, ANSM in France, BfArM in Germany), and expose the company to massive litigation. Manufacturing process data: proprietary formulation parameters, process recipes, quality control specifications, and yield optimisation data for 180+ products. The company's plasma fractionation processes (separating plasma into therapeutic proteins — albumin, immunoglobulins, clotting factors) were refined over 35 years of manufacturing experience. These processes represented hundreds of millions in R&D investment and were protected as trade secrets. A competitor accessing manufacturing process data could replicate products that took decades to develop. Pharmacovigilance data: adverse event reports, safety signals, and periodic safety update reports (PSURs) for all marketed products. Pharmacovigilance data included patient-identifiable health information and was regulated by EMA and national authorities — any data breach could trigger regulatory action and public safety concerns. Pre-patent research data: drug discovery data, molecular screening results, and formulation research preceding patent filing. Premature disclosure of this data could invalidate patent applications worth billions in future revenue. (2) Manufacturing quality AI — GMP requirements. The company's 12 manufacturing plants operated under EU GMP (Good Manufacturing Practice) requirements, inspected by AEMPS (Spain), ANSM (France), Regierungsprasidium (Germany), and INFARMED (Portugal). GMP required: complete traceability of manufacturing decisions (if AI influenced any manufacturing parameter, the AI's input, reasoning, and human decision must be documented), validated computer systems (EU Annex 11 requiring that computer systems affecting product quality are validated — AI systems falling under this requirement), and data integrity (ALCOA+ principles — Attributable, Legible, Contemporaneous, Original, Accurate — applied to AI-processed data and AI-generated recommendations). Cloud AI APIs could not satisfy GMP computer system validation requirements because the company had no control over model versioning, could not validate the system according to EU Annex 11, and could not guarantee data integrity end-to-end. Self-hosted AI with full model lifecycle control was the only GMP-compliant option. (3) Plasma fractionation optimisation — the company's most complex manufacturing challenge. Plasma fractionation involved separating human blood plasma into therapeutic proteins through a series of chemical and physical processes (cold ethanol fractionation, chromatography, filtration, viral inactivation). Each step had 40-60 controllable parameters (temperature, pH, ethanol concentration, centrifugation speed, buffer composition, incubation time) that affected yield, purity, and product quality. Optimising these parameters could: increase yield (extracting more therapeutic protein from the same volume of plasma — each percentage point of yield improvement worth approximately EUR 8M annually given the company's plasma volume), improve purity (reducing contaminants and improving product safety profiles), reduce batch failures (failed batches representing EUR 200K-500K in lost material plus production schedule disruption), and extend product shelf life (optimised formulation parameters affecting product stability). The company had 35 years of manufacturing batch records — 420,000 batch records with complete process parameter data and quality outcomes. AI could analyse these records to identify optimal parameter combinations, predict batch outcomes, and recommend adjustments. But this data — the company's complete manufacturing knowledge — could never leave company-controlled infrastructure. (4) Drug development acceleration — the company's R&D pipeline included 14 development-stage products. AI could accelerate: formulation screening (testing drug-excipient combinations computationally rather than through physical experimentation), stability prediction (predicting product stability under various storage conditions from molecular and formulation properties), clinical trial optimisation (identifying optimal patient populations, predicting enrollment challenges, and analysing interim results), and literature analysis (processing 12,000+ scientific papers annually to identify relevant research findings, competitive intelligence, and regulatory trends). R&D data was pre-competitive intelligence — any external exposure could compromise patent strategy and competitive positioning. (5) Pharmacovigilance signal detection — the company monitored safety data for 180+ marketed products across 90 countries. Current pharmacovigilance processes relied on manual case processing and statistical signal detection (proportional reporting ratio, empirical Bayesian methods). AI could enhance: case processing automation (extracting adverse event information from medical literature, spontaneous reports, and clinical trial data), signal detection (identifying safety patterns from narrative case descriptions that statistical methods missed), and PSUR preparation (automated periodic safety update report generation from accumulated case data). Pharmacovigilance was regulated by EMA and national authorities — AI processing of pharmacovigilance data required validated systems with complete audit trails.

We deployed a self-hosted AI platform over 14 weeks — GMP-validated, EU Annex 11-compliant, and purpose-built for pharmaceutical operations. (1) Barcelona-resident pharmaceutical-grade infrastructure. Hardware: 24 NVIDIA H100 80GB GPUs across 3 server nodes, deployed in the company's Barcelona data centre (a validated pharmaceutical computing facility meeting EU GMP Annex 11 requirements — controlled access, environmental monitoring, UPS-protected, and equipped with validated backup procedures). Validation: the AI platform validated as a GMP computerised system following EU Annex 11 and GAMP 5 methodology. Validation documentation including: User Requirements Specification (URS), Functional Specification (FS), Design Qualification (DQ), Installation Qualification (IQ), Operational Qualification (OQ), and Performance Qualification (PQ). Every model deployment, update, and configuration change documented in the company's change control system — meeting GMP requirements for validated system management. Data integrity: all AI inputs and outputs meeting ALCOA+ requirements — Attributable (linked to the requesting user and system), Legible (human-readable outputs), Contemporaneous (timestamped at creation), Original (tamper-evident storage), and Accurate (validated against known-good test cases). (2) Manufacturing quality AI: optimising 35 years of plasma fractionation. Batch record analysis: Llama 3 405B fine-tuned on 420,000 historical batch records (35 years of manufacturing data). The model learning correlations between process parameters and quality outcomes that traditional statistical analysis had missed — particularly multi-parameter interactions (the effect of pH adjustment timing combined with specific ethanol concentrations at specific temperatures) that were invisible to individual parameter analysis. Process parameter optimisation: the AI providing recommendations for process parameter adjustments — not controlling manufacturing equipment directly (GMP requiring human decision-making for production parameters) but presenting optimised parameter sets to manufacturing scientists with predicted outcomes and confidence intervals. For each batch, the AI recommended: optimal parameter targets (temperature setpoints, pH targets, centrifugation parameters), critical process parameters to monitor most closely (based on the specific plasma lot characteristics), and predicted yield and quality outcomes (enabling manufacturing scientists to set appropriate in-process testing schedules). Batch outcome prediction: before each batch began, the AI predicting: expected yield (protein recovery per litre of plasma), potential quality issues (based on input plasma characteristics and planned process parameters), and risk factors (parameters where the specific batch was most likely to deviate from specification). Quality deviation investigation: when batches deviated from specification, the AI analysing all available data to identify root cause — comparing the deviation batch's parameters against historical batches with similar outcomes, identifying the parameter deviations most likely responsible. Deviation investigation time: from 5 days average to 8 hours (AI providing root cause hypothesis within 2 hours, manufacturing scientists validating within 6 hours). (3) R&D AI: accelerating drug development. Formulation screening: AI analysing the company's formulation database (4,200 historical formulations with stability and performance data) to predict: drug-excipient compatibility for new molecules, optimal formulation composition (vehicle, preservative, buffer, tonicity agent), and predicted stability profile (shelf life estimation based on molecular properties and formulation composition). Formulation screening reducing physical experiments by 62 percent — the AI narrowing candidate formulations from 200+ to 20-30 for physical testing, dramatically reducing development time and cost. Literature intelligence: Llama 3 70B processing scientific literature — 12,000+ papers annually across the company's therapeutic areas (plasma therapies, hospital injectables, dermatology). The AI extracting: competitive intelligence (competitor pipeline updates, regulatory submissions, clinical trial results), scientific advances (new purification techniques, novel excipients, stability enhancement methods), and safety signals (published adverse events relevant to the company's products — feeding into pharmacovigilance). Literature processing: from 3-week review cycle (team of 8 scientists) to continuous AI processing with weekly synthesised reports reviewed by scientific team in 4 hours. Clinical trial analysis: AI processing interim clinical trial data — identifying enrollment trends, protocol deviations, and early efficacy signals. The model understanding clinical trial terminology and regulatory reporting requirements (ICH E6 R2, EU CTR 536/2014). (4) Pharmacovigilance AI: safety signal detection. Case processing: AI processing incoming adverse event reports (ICSRs — Individual Case Safety Reports) from multiple sources: spontaneous reports from healthcare professionals and patients, literature case reports, clinical trial adverse events, and post-marketing study reports. The AI extracting MedDRA-coded events, assessing causality, and populating the company's safety database — reducing case processing time from 45 minutes to 12 minutes per case while maintaining data quality validated through parallel processing (AI and human processing 500 cases independently, comparing outputs). Signal detection: LLM-enhanced signal detection analysing narrative case descriptions — identifying safety patterns from free-text descriptions that statistical methods (which operated on coded data only) missed. The AI identifying 8 potential safety signals in the first 12 months that traditional statistical methods had not detected — 3 of which were subsequently confirmed as valid signals requiring regulatory action. PSUR automation: AI generating draft periodic safety update reports — compiling case data, benefit-risk assessment summaries, and regulatory context. PSUR drafting time: from 6 weeks to 8 days (AI generating comprehensive draft in 2 days, medical writers editing and finalising in 6 days). (5) GMP model governance: pharmaceutical-grade AI management. Validated system lifecycle: every aspect of the AI platform managed through pharmaceutical validation lifecycle — requirements, specifications, qualifications, ongoing monitoring, periodic review, and decommissioning procedures. Change control: any model update, parameter change, or configuration modification processed through the company's pharmaceutical change control system — risk assessment, impact analysis, testing, approval, and implementation documentation. Audit trail: complete audit trail meeting AEMPS inspection requirements — every AI interaction logged with: user identification, timestamp, input data reference, AI output, human decision (accept/modify/reject), and rationale. AEMPS inspection readiness: documentation organised for regulatory inspection — the company able to demonstrate to AEMPS inspectors exactly how AI was used in manufacturing decisions, what controls were in place, and how the system was validated.

Results over 12 months post-deployment. Manufacturing: plasma fractionation yield improvement: 4.2 percent average yield increase across all protein products. At the company's plasma processing volume, this represented EUR 33.6M in annual additional product value (more therapeutic protein from the same plasma volume). Batch failure rate: from 2.8 percent to 1.1 percent (61 percent reduction). Each prevented failure saving EUR 200K-500K in materials and production capacity — estimated EUR 8.4M annual savings. Quality deviation investigation: from 5 days to 8 hours average (84 percent reduction) — faster root cause identification enabling faster corrective action and reduced production impact. Manufacturing science team reporting that AI "fundamentally changed how we optimise processes" — identifying multi-parameter interactions that decades of traditional analysis had missed. R&D: formulation screening efficiency: 62 percent reduction in physical experiments (from 200+ to 75 average candidate formulations per development programme). Estimated 4-month acceleration in formulation development timelines. Literature processing: from 3-week manual review cycle to continuous AI processing with weekly 4-hour review sessions. Scientific team identifying 14 competitively significant research findings through AI literature analysis that manual review would have missed or identified months later. Clinical trial analysis: interim data processing enabling 2 protocol amendments that improved enrollment and reduced study duration by estimated 3 months each. Pharmacovigilance: case processing: 8,400 ICSRs processed annually. Processing time: from 45 to 12 minutes per case (73 percent reduction). Data quality: 98.2 percent agreement between AI and human parallel processing. Signal detection: 8 potential safety signals identified by AI (3 confirmed — traditional methods detected 0 of these 3 within the same timeframe). PSUR preparation: from 6 weeks to 8 days (81 percent reduction). Regulatory submission timeliness: 100 percent on-time submission rate (versus 88 percent previously — late PSURs having triggered AEMPS corrective action requests). GMP compliance: EU Annex 11 validation completed and maintained through change control. AEMPS inspection (conducted 8 months post-deployment): zero findings related to AI systems. The AEMPS inspector commenting that the company's AI governance was "exemplary" and requesting permission to share the approach as a reference for other pharmaceutical companies. Cost: infrastructure investment: EUR 2.8M (24 H100 GPUs, servers, networking, validated computing environment). Annual operating cost: EUR 680K (including ongoing validation maintenance). Annual value: yield improvement EUR 33.6M, batch failure reduction EUR 8.4M, R&D acceleration (estimated) EUR 12M (faster time-to-market for pipeline products), pharmacovigilance efficiency EUR 2.4M, deviation investigation efficiency EUR 1.8M. Total annual value: EUR 58.2M. ROI: 16.7x in first year.