Deep dives
Problem, context, constraints, architecture, decisions, trade-offs, and outcome across professional and personal engineering work.
Personal project: treating retrieval infrastructure as a lifecycle problem, not a demo
A self-initiated, local-first RAG backend built to explore what production-grade retrieval infrastructure actually requires beyond a working demo: lifecycle management, multilingual retrieval, and framework independence.
Personal project: treating AI agents as an SDLC, not a code-completion tool
A self-initiated project exploring what happens when a requirement flows through a pipeline of specialized AI agents (planning, implementation, testing, review) end to end, instead of a single assistant writing code from a prompt.
Delivering a defined multi-boundary configuration flow reliably, not redesigning it
Building a React/Node configuration dashboard that lets application settings be updated live in Kubernetes, by loading a secret from HashiCorp Vault, running an AWS Lambda, updating a Kubernetes ConfigMap, and rerunning the pipeline.
Validation, defensive state handling, and scheduled cleanup in a live Java payments platform
Hardening validation and correctness and automating lifecycle cleanup in Intuit's QuickBooks Bill Pay B2B payments service: batch-validation and payment-state-aware abort fixes, a runtime-refreshable Payment Instruction expiration TTL, a ShedLock-coordinated scheduled stale-Payment-Instruction cleanup job built on a self-designed sparse, sharded DynamoDB GSI with authoritative-state revalidation before every abort, and a failed-credit-card read-path filter.
Independent architecture ownership, validated after the fact
Designing and owning a centralized Kafka routing layer for a banking platform, independently, and having the design validated by the principal architect after the fact rather than before.
Incremental legacy modernization with controlled blast radius, not a rewrite
Migrating a 15-year-old legacy tax scripting pipeline onto a modern internal execution engine, provider by provider, using shadow validation to keep blast radius small instead of cutting the whole system over at once.