John Terry

Developed backend components for conversational product discovery and comparison, sustaining 200 requests per second under load testing. Integrated relevant product information into response generation workflows to help customers compare features and understand available options. Implemented conversational context handling to support follow-up questions and preserve meaning in multi-turn shopping interactions. Added validation for missing product attributes, conflicting source information, and responses unsupported by available evidence. Built evaluation scenarios across five assessment areas: shopping intent, product relevance, factual accuracy, incomplete information, and response latency. Improved service resilience through timeouts and dependency error handling, reducing customer-facing request failures from 1.5% to 0.9%.

Developed enterprise assistant integrations using Amazon Q Business, supporting information discovery for 500 monthly active users. Built document synchronization workflows for content and permission updates, reducing p95 synchronization lag from 60 minutes to 15 minutes. Integrated identity and document-permission mappings so retrieval respected access restrictions from source systems. Developed API integrations for supported business actions with input validation, authorization checks, and recoverable error handling. Built test scenarios covering retrieval relevance, source attribution, stale content, and unauthorized access to indexed information. Integrated Amazon Q Developer into three engineering workflows (code explanation, test development, and troubleshooting) with engineering review of generated suggestions. Added monitoring for connector health, synchronization delays, ingestion errors, and assistant failures to support operational investigations.

Developed backend services integrating Amazon Bedrock model APIs across three generative AI workflows: summarization, question answering, and document processing. Built retrieval-augmented generation pipelines to provide source-grounded context, reducing p95 retrieval latency from 800ms to 520ms. Implemented reusable integration components for prompt construction, response parsing, request validation, and error handling, and added timeouts, bounded retries, concurrency controls, and fallback behavior to improve reliability when downstream services or model endpoints were unavailable. Added authorization checks and scoped retrieval to protect sensitive information throughout the request lifecycle. Created automated evaluation workflows for answer relevance and factual grounding, reducing unsupported answers from 12% to 8% on a fixed evaluation set. Added operational metrics and dashboards for model errors, token usage, latency, and resource consumption.