Nishtha Gupta

A verification service where I owned the network and infrastructure layer, then kept extending it after the team build ended.

• Designed AWS networking and container infrastructure across a multi-AZ VPC, ALB, ECS Fargate, and port-scoped security groups, with the full stack provisioned and torn down via 10 modular Terraform components.
• Decoupled claim processing from the verification response path using SQS with a dead-letter queue (maxReceiveCount=3, 14-day retention), so verification stays available independently of claims-service failures.
• Reduced p99 verification latency 20–50× — 80–150ms down to 2–10ms — with an ElastiCache Redis cache-aside layer, while excluding WINNER outcomes from caching to prevent stale claim state across agents.
• Benchmarked with a purpose-built script and custom CloudWatch metrics (CacheHit, CacheMiss, VerificationLatency) before and after Redis, validating the cache across both cold and warm request paths.
• Evaluated database-per-service decomposition and retained a shared RDS instance by design, with SQS payloads carrying enough context for the claims service to operate independently of the verification data tier.

A streaming climate-data pipeline that ingests, persists, and queries records through event-driven workflows.

• Orchestrated the pipeline across Kinesis, DynamoDB, S3, Athena, and EventBridge.
• Designed service access patterns around stream buffering, key-value lookup, object storage, and columnar analytics — mapping each workload to Kinesis, DynamoDB, S3, and Athena respectively.
• Separated ingestion, storage, and analytical workloads through event-driven components, so each can be operated and scaled independently.

A retrieval-augmented agricultural advisory agent that generates attributed crop recommendations grounded in retrieved evidence rather than unsourced text.

• Architected the agent in Python to retrieve region-specific agricultural sources and attribute every recommendation to them.
• Deployed the RAG pipeline on AWS with application and retrieval observability, enabling end-to-end diagnosis of latency across retrieval and inference stages separately.
• Formulated a fixed evaluation set for repeatable retrieval-quality testing, making agent behaviour comparable across iterations instead of judged by inspection.