Selected system work

A few systems, described by what changed.

Real work, framed by problem and outcome rather than client names. Some details are generalized to respect confidentiality.

Search · Streaming

240m → 6m~40× fasterNo backfill stalls

OpenSearch ingestion pipeline rebuild

Problem: The existing DMS + Kinesis pipeline was slow, and large indexing runs blocked new data from being indexed — a hard ceiling on throughput.
Approach: Re-architected the path to Debezium → MSK (Kafka) → OpenSearch Ingestion, decoupling backfills from live writes and parallelizing the heavy lifting.
Outcome: Indexing 1M records went from 240 minutes to 6 minutes (~40× faster), and large runs stopped starving new indexing.

Cloud · IaC

$30k → $22.5k / mo~$90k / yr savedRepeatable IaC

Problem: Production infrastructure was managed by hand in the console — hard to review, easy to drift, and expensive to run.
Approach: Migrated compute, messaging, data, ingress, delivery, security, and networking to AWS CDK, then modernized selected workloads through architecture changes, usage-based right-sizing, opt-in cost controls, and Graviton moves where they fit.
Outcome: Infrastructure changes became repeatable and code-reviewed, while monthly spend dropped from $30k to $22.5k (~$90k/year) after CDK migration, usage-based container right-sizing, Graviton moves, and opt-in AWS cost controls.

Backend · Latency

~95% fasterArchitectural redesignHot-path tuning

Problem: A customer-facing Inbox had slow response times that hurt day-to-day usability, especially under load.
Approach: Redesigned the architecture and optimized the hot paths end to end, rather than patching around the symptoms.
Outcome: Cut response times by roughly 95%, making the Inbox feel near-instant again.

Backend · Migration

PHP → Node.jsCritical-path rewriteLower maintenance

Problem: The Sorter — the most critical component routing messages through the system — was an aging PHP service that was hard to evolve and maintain.
Approach: Rebuilt it in Node.js / TypeScript with a cleaner design, modernizing the critical path while preserving behavior.
Outcome: Improved the performance and maintainability of the system's single most vital component.

Database · Scale

1B+ rows in one database tablePartitioned by dateSafer reporting load

Problem: A production reporting table had grown past a billion rows, and query patterns that once worked were degrading under real traffic.
Approach: Combined denormalization, read replicas, cursor pagination, targeted index hints, short-TTL caching, monthly date partitioning, enforced date filters, archiving, rate limits, and review guardrails for future queries.
Outcome: Kept reporting queries efficient on the active dataset, reduced production load, and made future access patterns safer to evolve.

SaaS · Multi-tenancy

Multi-tenant SaaSMillions of customersLegacy migration

Problem: Large insurance brokerages needed to manage millions of customers across isolated tenants on a single platform.
Approach: Helped build a multi-tenant SaaS product and led a migration that moved data off a legacy single-tenant system onto it.
Outcome: Adopted by several of the world's largest insurance brokerages to manage millions of customers more efficiently.