I build and ship zero-to-one AI platforms, from the first commit to the team that outlasts me.

Problem

Clients needed agentic workflows shipped quickly, but the usual trade-off was brutal: move fast and accept brittle, non-deterministic pipelines, or move slow to earn durability and reliability. The mandate was to get both: production-grade reliability without sacrificing the speed clients were paying for.

Architecture

AgencyOS is a workflow-orchestration layer built on top of LittleHorse, a durable-execution engine in the same problem space as Temporal. The design treats durable execution as a first-class concern: long-running, multi-step agentic flows are modeled as durable workflows so that state survives restarts, retries are deterministic, and partial progress is never lost.

On top of that foundation I added durability primitives and an opinionated set of conventions that let engineers author reliable workflows without re-learning the distributed-systems failure modes each time.

Outcome

A production framework now powering multiple concurrent client engagements. The durable foundation produced qualitative gains in both delivery speed and reliability: workflows that would previously have been one-off, fragile scripts became reusable, observable, and resilient by default.

Problem

Build a music-discovery and social-matching product where recommendations are graph-native, surfacing both the right tracks and the right people from how a listener actually behaves, not from shallow genre tags.

Architecture

I designed the Neo4j graph architecture that backs recommendations, deployed on Kubernetes for elastic scale. On top of it I built Music Match, a taste-embedding microservice in Python that matches user to user from listening behavior, and a production inference pipeline that generates the recommendation embeddings serving the graph.

Backend services spanned NestJS and FastAPI, with the full footprint provisioned as code via Terraform across AWS.

Outcome

Needle launched to Top 10 on the App Store on day one.

The Christmas launch

Launch day landed over the holidays, and traffic immediately outran the infrastructure while the team was scattered across the country for travel. I scaled the database and backend mid-flight, coordinated the team remotely, and stayed on-call through the holiday. The platform stabilized under load and the Top 10 ranking held. It was the clearest test of owning a system end to end: technical depth under pressure, with no one else able to take the keyboard.

Problem

The designer-to-developer handoff loses time and fidelity at every iteration: designers ship pixel-perfect Figma files, developers hand-translate them into code, and the two drift apart the moment either side changes. The goal was to automate that translation: turn a design file directly into clean, responsive Flutter code so design and development could move continuously.

Architecture

As CTO I led Parabeac-Core, the open-source conversion engine, written in Dart. It pulls a design straight from the Figma API and lowers it through a staged export pipeline emitting idiomatic Flutter.

Around the engine I also helped build the SaaS web app (frontend and backend) that wrapped the core into a product designers and teams could actually use.

Outcome

Parabeac-Core shipped as an open-source project with a real community behind it: ~600 GitHub stars across 30+ releases, and became the foundation that shaped how I think about closing the gap between design and engineering.

github.com/Parabeac/parabeac_core →