Design deliverables for the quote-ingestion pipeline (Gmail → extracted quote in RDS) and the RFQ matching + submission tail. Each card renders a diagram or spec; yellow = AI-bearing throughout.
Gmail poll → fetch → classify → extract & persist to RDS.
The logical journey of a quote (stages 1–4), plus the decomposed Stage-4 extraction sub-pipeline.
Private VPC, scheduled-poll notifications, SQS between stages, Bedrock, and the CloudWatch → SNS → PagerDuty alerting.
Prompt-engineering over Bedrock, versioned templates, schema-constrained output, embeddings, and evaluation.
ExtractedQuote → matched to a customer RFQ → submitted to the core system.
Matching + submission (stages 5–6), plus the decomposed retrieve-then-rerank matching sub-pipeline.
Tighter posture — no NAT, no internet egress; RFQ read-model sync, Fargate matching over pgvector, private submit.
Three-model retrieve-then-rerank stack: embedding recall → cross-encoder precision → LLM disambiguation.
Shared data model and the API contract across both stages.