A few selected projects. Each with problem, solution, stack, and lesson learned.
Open-source validator for typed agent plan artifacts: the boundary object between agents that plan and the runtime that commits. Four static checks (pivot ordering, class coherence, mandate limits, undo TTLs) reject structurally unsafe plans before a single side effect runs, including plans whose rollback windows cannot survive a human approval step. MIT-licensed and on npm.
Open specification proposal that turns platform compromises into recorded, queryable data. Developers declare desired outcomes; the platform matches them against human-approved capabilities and records every negotiation — granted, degraded, or rejected — with the reasoning and approver, instead of letting it evaporate into a Slack thread.
Open-source Rust framework for CDC (Change Data Capture) and data replication, built for real-time data pipelines. Captures MongoDB change streams and replicates them to destinations like S3 (JSON, CSV, Parquet, Avro), with a Redis-backed distributed state store and production throughput from 10K to 100K events/sec.
Migrated a mission-critical weather alert notification system from on-premise infrastructure to a fully serverless AWS architecture, enabling immediate delivery of safety alerts during high-traffic bursts. Designed high-throughput, asynchronous microservices for multi-channel notifications.
Led the architectural integration of AI-driven workflows for automated vehicle damage assessment, partnering with product and business teams to turn a complex automation problem into a user-facing capability.
Built event-driven data pipelines with Apache Kafka and Flink to calculate real-time driver risk scores, a pay-per-use telematics insurance platform. Highly available REST APIs in Java and Kotlin on AWS for large-scale, real-time processing.