Laiba Shahab

Summary

What I did

• Engineered hybrid RAG combining pgvector similarity search, PostgreSQL full-text search and reciprocal rank fusion, grounding AI answers in authorized governance, risk and compliance evidence with citations.
• Centralized AWS Bedrock inference and streaming with structured-output validation, bounded repair, tenant token budgets and usage accounting; implemented persona-based golden-answer evaluation.
• Implemented Cognito JWT authentication and enforced PostgreSQL row-level security for tenant isolation; built event-driven ingestion and background processing with ECS/Fargate, S3, SQS and EventBridge.
• Implemented a hybrid retrieval pipeline using pgvector semantic search and PostgreSQL full-text search with Reciprocal Rank Fusion.
• Developed an LLM orchestration layer to manage retries, timeouts, idempotency, and structured-output validation.

Results

• Centralized AWS Bedrock calls behind an LLM orchestration layer to handle token estimation, pricing, tenant-level usage limits, and usage reconciliation.
• Implemented tenant token budgets and usage accounting to prevent unbounded model consumption and attribute usage back to tenants.