Software Development Engineer
Amazon · Austin, TX
Led a multi-region payments migration and data-residency compliance for live traffic (~175K transactions, ~$13B USD/yr) with zero customer impact. Built parallel migration workers across DynamoDB partitions with zero duplicate seller payouts, re-engineered routing behind secure endpoints, and executed a seamless full-traffic cutover.
- —Designed an active-passive failover system with sub-5-minute recovery and near-zero data loss, sustaining 99.95% availability.
- —Instrumented 10+ Java/Spring Boot microservices for end-to-end distributed tracing, cutting MTTR ~30% (from ~40 min to ~28 min).
- —Decomposed a tightly-coupled payments monolith into independently deployable microservices, improving deploy flexibility.
- —Built an agent-based automation platform (Python, A2A + MCP) that auto-runs ~30% of on-call runbooks and raised straight-through processing ~12%.
- —Built an LLM-powered test-triage tool that reproduces failing tests, finds root cause, and drafts fixes; adopted by the core platform team.
- —Built a full-stack reconciliation console (React/TypeScript, Java REST APIs) that tripled bulk payment-reconciliation throughput.
- —Cut downstream service calls ~20% and peak infrastructure cost ~15% via ingress-level validation and adaptive throttling.
- —Built a shadow-execution framework that diffed new vs. legacy payment paths in production, safely retiring ~40% of legacy code.
- —Rebuilt the payment-processor test simulator into a plugin-based framework, adopted across teams to raise integration coverage.
- —Held on-call for a Tier-1 payments service — detection, mitigation, and root-cause fixes that eliminated recurring incidents.
- —Hardened CI/CD with canary and staged rollouts and automated rollback on health-check regression, shortening deploy cycles.
- —Built peak-volume load and performance tests that replayed production traffic to surface bottlenecks before launch.
- —Optimized a hot database access path with query/index tuning and targeted caching, cutting p99 latency from ~900 ms to ~500 ms on a high-traffic endpoint.
- —Reduced false-positive alerts by retuning alarm thresholds and consolidating redundant monitors, surfacing real incidents faster.
- —Cut cloud spend ~18% by right-sizing ECS tasks and Lambda memory and moving high-volume DynamoDB tables to on-demand with TTL-based expiry.
- —Reduced Lambda cold-start latency on payment endpoints ~40% (p99) via provisioned concurrency and slimmer deployment payloads.
- —Secured 26 third-party settlement integrations with tokenization, fail-closed controls, and immutable audit logs on least-privilege roles.
JavaSpring BootPythonGoAWSDynamoDBReact/TSA2AMCP