Developed an AI caregiving assistant using AI SDK v6, combining RAG over managed content, web search tools, long-term memory and conversation persistence to deliver personalized support for 1M+ caregivers.
Built Agent QA - a deterministic testing framework for production AI agents, enabling YAML-based multi-turn scenario testing across tool calls, database mutations, conversation flow, token usage and hallucinations. Added multi-run reliability analysis, A/B testing, LLM-as-a-judge, per-agent token attribution and a Langfuse integration to improve observability, reliability and cost efficiency.
Developed AgentKit - a Typescript framework for building real-time durable agents that can pause/resume reliably via checkpointing, deterministic routing, persisted state and MCP tooling - primarily designed for long-running background agents and deployable via Inngest's durable execution platform across all major cloud providers.
Integrated long-term memory and conversation persistence via a database-agnostic history adapter to resume agent runs reliably across executions. Shipped human-in-the-loop workflow capabilities - enabling agents to pause for user input/approval and continue execution inside durable functions.
Shipped a React SDK for durable, real-time streaming and thread management - enabling rapid development of production-grade agent UIs. Extended developer tooling for fine-grained agent observability including run tracing, token-usage inspection and reruns with modified prompts at any step in an agentic workflow.
Led 3 teams (18 engineers) building an interoperability platform integrating EHRs (Epic, Cerner, AllScripts), referral networks, HIEs and payers to support closed-loop referrals and population health reporting across NY and MA Medicaid 1115 initiatives.
Built a self-service data ingestion and analytics platform supporting REST, streaming and CSV-based onboarding, with normalization, validation and reporting workflows. Designed multi-tenant governance and least-privilege access controls for securely sharing data across healthcare organizations, backed by AWS, Kafka and Java/Quarkus.
Architected React-based embeddable micro-frontends and EHR integrations with data visualization for population health reporting and care coordination use cases.
Grew an audience of 15k followers on X - sharing insights from hundreds of research papers on the practical application of AI workflows, agentic systems, long-term memory, information retrieval and knowledge graphs.
Shipped BrowserTools (7.3k Github stars) - an open-source Typescript MCP server and Chrome extension that streams browser telemetry into AI coding agents for automated debugging/validation. Became the most widely adopted MCP server by Cursor users throughout Q1/Q2 of 2025 and featured by A16z.
Designed and implemented an in-house, Vercel-like serverless deployment platform for NextJS on AWS (S3, Cloudfront, WAF, Lambda/API Gateway) - automated via Jenkins pipelines and Groovy-based CI/CD adopted by 300+ engineers to enable rapid and secure iteration of consumer healthcare web apps.
Launched member-facing NextJS applications, including benefits explorer, provider directory and scheduling tools used by 1M+ users (members, prospects, employees) - improving access to coverage and care information.
Developed landing pages and analytics dashboards using React/Redux & NodeJS/Express deployed on AWS EC2 and S3/Cloudfront. Improved conversions by automating email campaign flows and surfacing KPIs (traffic, revenue, funnel metrics) into both customer facing and internal dashboards.
Built and maintained a client-facing marketing analytics and content management platform using React, Redux, Firebase and Python - visualizing global KPIs across distributed content, influencer and advertising campaigns with integrated Google Analytics and SimilarWeb data.
Implemented a serverless ETL-style API on AWS Lambda and API Gateway (Python/Flask, Serverless Framework) to aggregate and normalize data from Google Analytics and SimilarWeb.
Co-authored and edited whitepapers for blockchain-based initial coin offerings (ICOs), translating technical concepts, token economics and product strategy into investor-facing documentation.
Built a PLM system and file sharing application (React, Node/Express, MongoDB) for bi-directional data sync between AWS and China-based cloud providers (60% cost reduction and 18x faster downloads).
Vibe coding produces technical debt at scale and decouples developers from understanding how their systems actually work. Learn how to structure a monorepo with hexagonal architecture, documentation layers, comprehensive testing, and custom tooling to make AI coding agents genuinely effective.
LLM evaluations are not the end-all-be-all of AI quality assurance. Here's how I built Agent QA - a deterministic agent testing framework that makes AI self-testing possible.
A contrarian take on why reaching for your mouse might actually make you a better programmer. Vim enthusiasts, look away.
A deep dive into ModernBERT—the next evolution of encoder-only transformers, designed for longer contexts, better efficiency, and state-of-the-art performance in retrieval and classification.
Always interested in new opportunities, collaborations, and conversations about technology and design.