Manufacturing Celestara

AI-Powered Yield & OEE Optimisation

Yield losses in pharmaceutical manufacturing are rarely caused by a single variable. They emerge from the interaction of equipment performance, environmental conditions, material variability, and operator decisions across thousands of batch records. Yet most manufacturers still investigate yield gaps manually, batch by batch, with no systematic way to connect the signals. Celestara deploys specialised AI agents that sit on top of your MES and logbook data to identify patterns behind yield losses, recommend corrective actions that operators and engineers can understand and act on, and continuously track whether those recommendations actually improve outcomes—completing the loop that manual processes never close.

Key Highlights

01 Real-time visibility into yield trends during production
02 AI-identified root causes with human-explainable reasoning
03 Closed-loop tracking of recommendation effectiveness
04 Multi-site OEE benchmarking and comparison

Regulatory Context

What FDA Inspectors Cite

21 CFR 211.103 Yield Calculation
1 / 6

Batch yield calculations exclude visual inspection defects without scientific justification

The firm inappropriately calculated batch yields by leaving out all visual inspection defects, which potentially accounted for significant loss of the batch. There was no scientifically sound justification for excluding inherent particulate rejects from percent defect or yield calculations (actual versus theoretical yield).

Pine Pharmaceuticals, LLC · 2024-09-30
21 CFR 211.103 Phase Yield Tracking
2 / 6

Failure to calculate theoretical and actual batch yield at each production phase for sterile drugs

The firm failed to adequately calculate the theoretical and actual batch yield steps at each phase of the production process for all sterile drug products. Calculated yield steps were not inclusive of all products produced during the manufacturing run. Production proceeded to the next step when actual yield failed to meet the validated range.

SCA Pharmaceuticals, LLC · 2024-12-20
21 CFR 211.188 Yield Discrepancy
3 / 6

Raw material quantities in batch records do not support actual API yields observed

Batch production records documented output quantities that did not match actual material found in warehouses. Extra drums of API were discovered with no traceability to manufacturing history, raw materials, or process conditions. The firm could not reconcile documented yields against physical inventory, making traceability of disposition impossible.

Hetero Labs Limited, Unit-IX · 2025-09-26
21 CFR 211.68 Equipment Performance
4 / 6

Equipment maintenance failures cause numerous BMS excursions affecting production quality

Review of BMS data for 2022–2024 revealed too numerous to count excursions of differential pressure, temperature, and humidity throughout production areas. At least 40 work orders for aseptic equipment repairs extended over 100 days. Recurring issues including leak test failures caused process time extensions and batch rejections. Preventive maintenance records failed to identify instruments used for verification.

Biocon Biologics Limited · 2024-07-26
21 CFR 211.188 Equipment Traceability
5 / 6

Batch production records fail to include identity of major equipment and lines used

There was no identification of critical equipment throughout batch production records. Batch records for drug product and finished drug product batches failed to include equipment identification, making it impossible to correlate equipment performance with yield outcomes or conduct meaningful OEE analysis.

Baxalta US Inc. · 2025-09-18
21 CFR 211.188 Production Records
6 / 6

Manufacturing calculations undocumented in batch records across four years of production

Critical calculations required during manufacturing were not documented on the batch records for all lots manufactured across multiple production lines from May 2020 to July 2024. Without documentation of these calculations, there is no audit trail for yield determination or variance investigation at any production phase.

Laboratorios Grifols, S.A. · 2024-07-26

The Problem

Why Yield Losses Go Undetected and Unresolved

Challenge 1 1 / 7

Unstructured Data from Logbooks and Manual Records

  • Critical yield-impacting events recorded as handwritten notes or unstructured text entries
  • Logbook data never enters the analytical pipeline—it stays in binders or siloed digital logs
  • Pattern recognition across logbook entries is impossible without manual review of thousands of records
Challenge 2 2 / 7

MES Data Silos Block Contextual Analysis

  • Batch yield figures sit in MES while equipment data lives in CMMS and environmental data in BMS
  • No automated correlation between a yield dip and the equipment, room conditions, or material lot that caused it
  • Investigators manually export data from 4–6 systems to build a single yield investigation timeline
Challenge 3 3 / 7

No Real-Time Visibility into Yield Trends

  • Yield calculations happen at batch close, not during production when corrective action is possible
  • Supervisors rely on end-of-shift summaries rather than live dashboards showing yield trajectories
  • Emerging yield trends across consecutive batches go unnoticed until they become significant losses
Challenge 4 4 / 7

Manual Root Cause Analysis for Yield Losses

  • Engineers spend days pulling data from MES, LIMS, ERP, and logbooks to investigate a single yield event
  • Investigations focus on the immediate batch rather than patterns across hundreds of similar batches
  • Root causes default to operator error or material variability because systemic factors are invisible
Challenge 5 5 / 7

No Feedback Loop on Recommendations

  • Process improvement suggestions live in deviation reports that are closed and never revisited
  • No mechanism to measure whether a parameter change actually improved yield on subsequent batches
  • The same recommendations are made repeatedly because no one tracks their effectiveness
Challenge 6 6 / 7

Multi-Site Yield Comparison Is Impossible

  • Each site captures yield data in different formats, at different granularity, using different MES configurations
  • A site achieving 95% yield on a product cannot systematically share what drives that performance with a site at 88%
  • Enterprise-level yield dashboards require months of manual data harmonisation
Challenge 7 7 / 7

Rising FDA Expectations on Process Monitoring

  • 483 observations cite failures to calculate yield at each production phase, not just at batch completion
  • Inspectors expect firms to detect and investigate yield discrepancies proactively, not reactively
  • Process validation lifecycle approach under FDA guidance requires ongoing monitoring with statistical justification

Yield Optimisation in Celestara

1

Unified Data Collection Across All Sources

Celestara continuously ingests data from MES batch records, digital logbooks, LIMS results, ERP material data, equipment sensors, and environmental monitoring systems. Unstructured logbook entries are parsed and structured by the Data Collection Agent, creating a unified production context that connects yield figures to the conditions that produced them.

Multi-Source Ingestion Logbook NLP Parsing Real-Time Streaming Schema Harmonisation
2

Real-Time Yield Tracking and Anomaly Detection

The Yield Analytics Agent monitors yield at every production phase in real time—not just at batch close. Statistical process control charts flag when yield trends deviate from historical norms. Supervisors see live dashboards that show yield trajectories during production, enabling intervention before a batch is lost.

Phase-Wise Yield Tracking SPC Charting Variance Detection Threshold Alerting
3

Multi-Variable Pattern Recognition

When a yield anomaly is detected, the Pattern Recognition Agent correlates the yield data with equipment performance, environmental conditions, material lot attributes, and operator logbook entries. It searches historical batches for similar patterns and identifies which variables most strongly predict yield outcomes—surfacing the systemic factors that manual investigation misses.

Multi-Variable Correlation Historical Matching Feature Importance Ranking Anomaly Clustering
4

Human-Explainable Recommendations

The Recommendation Engine translates pattern analysis into specific, actionable suggestions that engineers and operators can understand and evaluate. Each recommendation includes the data evidence behind it, the expected yield impact, and a confidence score. Recommendations are prioritised by potential yield recovery and feasibility of implementation.

Explainable Insight Generation Parameter Optimisation Impact Estimation Action Prioritisation
5

Closed-Loop Feedback and Efficacy Tracking

Every recommendation is tracked from issuance through implementation to outcome. The Feedback Loop Monitor Agent measures whether the recommendation was adopted, whether yield improved on subsequent batches, and by how much. Recommendations that prove effective are reinforced; those that do not are flagged for review. This completes the loop that manual processes never close—turning yield optimisation from a one-time investigation into a continuously improving system.

Implementation Tracking Before/After Yield Comparison Efficacy Scoring Continuous Learning

The Solution

How Leucine Solves This

Purpose-built tools that address both the FDA compliance gaps and the operational barriers to effective yield management and OEE optimisation.

Celestara Yield Intelligence

Unified Yield Visibility Across All Data Sources

Celestara ingests and correlates data from MES, logbooks, LIMS, equipment sensors, and environmental monitoring to provide a single view of yield performance. AI agents parse unstructured logbook entries, track yield at every production phase in real time, and detect anomalies before batches are lost—replacing post-hoc spreadsheet analysis with continuous, contextual yield intelligence.

  • Real-time yield tracking at every production phase with SPC charting
  • AI-powered parsing of unstructured logbook and manual record data
  • Multi-variable correlation linking yield to equipment, environment, and material factors
Celestara Recommendation Engine

Explainable AI Recommendations with Closed-Loop Tracking

The recommendation engine translates yield pattern analysis into specific, human-explainable actions prioritised by expected impact. Each recommendation includes supporting evidence, confidence scores, and feasibility assessment. Every recommendation is tracked from issuance through implementation to measured outcome—creating a closed feedback loop that continuously improves yield performance.

  • Human-explainable recommendations with data evidence and confidence scores
  • Automatic tracking of recommendation implementation and adoption
  • Before/after yield comparison to measure recommendation effectiveness
MES

Multi-Site Yield Standardisation and OEE Benchmarking

Deploy standardised yield calculation methodologies, equipment performance tracking, and OEE metrics across all manufacturing sites. Every facility captures yield data at the same granularity and in the same format, making cross-site comparison meaningful and enabling best-practice transfer from high-performing lines to those with yield gaps.

  • Standardised yield calculation at every production phase across all sites
  • OEE dashboards covering availability, performance, and quality rate per line
  • Cross-site yield benchmarking for identical products across facilities
FDA Tracker

Monitor Evolving FDA Expectations on Yield and Process Monitoring

Track how FDA cites yield calculation failures, production record deficiencies, and equipment performance gaps across the industry. Understand which yield-related deficiencies draw 483 observations and warning letters so your team can proactively align practices with current enforcement expectations.

  • Real-time tracking of 483 observations citing yield calculation and production record failures
  • Warning letter analysis for process monitoring and equipment qualification deficiencies
  • Benchmarking your yield documentation practices against industry enforcement trends

Results

Measurable Impact

Projected results from deploying Celestara for yield and OEE optimisation across pharmaceutical manufacturing operations

3.5%
Yield Improvement
Average absolute yield increase across production lines through AI-identified process optimisations
12%
OEE Increase
Improvement in Overall Equipment Effectiveness through visibility into availability, performance, and quality rate drivers
75%
Investigation Time
Reduction in time spent investigating yield deviations, from days of manual data gathering to hours with AI-assembled context
82%
Recommendation Effectiveness
AI recommendations that measurably improved yield on subsequent batches, verified through closed-loop tracking

Next Step

Get Started

Yield losses are not random—they follow patterns buried across your MES, logbooks, and equipment data. Celestara AI agents surface those patterns, recommend specific actions your teams can understand, and track whether those actions actually work. Stop investigating yield problems one batch at a time. Start seeing the system-wide picture.

Get Started
Exit