How Indegene Cut Manual Data Quality Work by 80% with Amazon Bedrock

About Customer:

Indegene is a global enterprise that partners with life sciences companies to accelerate their commercial, medical, regulatory and R&D operations. Headquartered in Bengaluru, the organisation combines deep domain expertise with advanced technology to help pharmaceutical, biotechnology and medical device companies navigate the complexities of the healthcare ecosystem.
Its solutions span digital transformation, omnichannel engagement, real-world evidence and regulatory compliance, supporting clients in improving operational efficiency and health outcomes. The work described here sits in the data layer beneath all of it: the data quality controls that regulated life sciences analytics depend on.

Industry:

Healthcare / HealthTech

Offering:

Gen AI/AIML

Business Challenges:

1. Data quality rules were written by hand. Data engineers and scientists authored SQL and PySpark validation checks manually for every dataset and every source, which made data quality a bottleneck ahead of any analytics or regulatory submission and consumed engineering capacity that should have gone to higher-value work.
2. Generic validation rules missed life sciences context. Off-the-shelf data quality checks do not understand pharmacovigilance terminology, regulatory field definitions or organisational policy, so rules were either too loose to catch real problems or too rigid to survive a schema change, and errors surfaced late as audit findings rather than early as failed checks.
3. Data sat across Amazon Redshift, relational databases, flat files and APIs. Each source was validated in its own silo with its own scripts, so the same logical rule was implemented several times in several dialects, and inconsistencies between sources went undetected until downstream reporting disagreed.
4. Subject matter expertise had nowhere to go. When a data expert corrected or improved a rule, that judgement lived in a ticket or a person’s head rather than in the system, so the same corrections were made repeatedly and rule quality never compounded.
5. Audit evidence was assembled manually. Life sciences regulation requires a demonstrable trail of what was checked, by whom, and when. Reconstructing that trail across scattered scripts and ad hoc executions was slow, and any gap in the record was an audit exposure.
6. Rule generation could not be reused at scale. Every new dataset restarted the process from zero, with no reuse of prior embeddings, queries or templates, so processing time and cost grew in step with the number of datasets rather than flattening as the library matured.

Solution Implemented:

1. Rapyder built a GenAI-powered data quality platform on AWS. Users start a DQ process from a central web application behind Amazon API Gateway; a classification AWS Lambda function calls an Amazon Bedrock Agent running Claude Sonnet 3.7 and Claude Sonnet 3.5 to identify the check type, and API Gateway then triggers the matching Lambda workflow for rule generation or SQL and PySpark query generation.
2. Grounded rule generation in Indegene’s own domain material using RAG. An Amazon Bedrock Knowledge Base powered by Claude Sonnet references structured reference PDFs held in Amazon S3, with the vector index in Amazon OpenSearch Service, so generated rules carry life sciences terminology, regulatory definitions and organisational policy rather than generic validation logic.
3. Added a dedicated validation step. Generated rules pass through a purpose-built Lambda function that uses Bedrock for context-aware review before anything reaches a user, and the resulting SQL and PySpark scripts are written to Amazon S3 for review, reuse and version control.
4. Unified execution across sources. Lambda orchestration runs the generated queries against Amazon Redshift, relational databases, flat files and APIs from a single control path, so one logical rule is expressed once and applied consistently everywhere.
5. Closed the loop with subject matter experts. Feedback captured in the UI is stored in Amazon DynamoDB, a feedback processing Lambda refines rule generation from those expert inputs, and updated rules and domain insights are folded back into the Bedrock Knowledge Base so later generations inherit the correction.
6. Made compliance a property of the platform. Enterprise identity management restricts rule access and approval to authorised roles such as Admin and SME, Amazon CloudWatch monitors executions and performance, and AWS CloudTrail maintains the audit log. Custom checks can be built against organisational policy and external compliance sources gathered by web scraping.
7. Engineered for reuse. Parallel processing and reuse of vector embeddings cut redundant work on new rule generation, and reusable queries and rule templates are stored in Amazon S3 as durable organisational knowledge.

Services Used:

• Amazon Bedrock
• Amazon Bedrock Agents
• Amazon Bedrock Knowledge Bases
• Claude Sonnet 3.7
• Claude Sonnet 3.5
• Amazon OpenSearch Service
• Amazon Redshift
• Amazon DynamoDB
• AWS Lambda
• Amazon API Gateway
• Amazon S3
• Amazon CloudWatch
• AWS CloudTrail
• AWS IAM
• Rapyder GenAI Advisory & Solution Design

Business Benefits:

1. Over 80% of routine data quality checks automated – Rule generation, validation and execution run without manual authoring, cutting operational errors and lifting data engineering team productivity by 30–40%.
2. Rule accuracy above 95% – Context-aware, domain-specific rules grounded in life sciences reference material reduce the risk of regulatory issues and audit findings compared with generic validation logic.
3. 2–3x improvement in data processing scalability – A single orchestration path across Redshift, relational databases and flat files lets the organisation absorb growing volumes of complex life sciences data without proportional engineering effort.
4. 20-25% gain in rule accuracy from the feedback loop – Expert corrections captured in the UI feed back into the Bedrock Knowledge Base, improving model performance and rule suggestions within the first few months of deployment.
5. 100% traceability for audits – Every rule execution, feedback event and system change is logged in Amazon DynamoDB and monitored through CloudWatch and CloudTrail, so compliance evidence is a query rather than a reconstruction exercise.
6. Reusable rules and embeddings across datasets – Parallel processing, reused vector embeddings and versioned SQL and PySpark templates in Amazon S3 remove redundant generation work and retain the rule library as organisational knowledge.

Is manual data quality work slowing down your regulated analytics?

Talk to Rapyder's GenAI and Data team to see how domain-aware DQ automation on Amazon Bedrock can cut manual effort and tighten audit readiness.