HACK DANCE

Opinionated local development orchestration for running multiple projects, branches, and worktrees at once without port conflicts. Stable local hosts, better logs, agent setup, and less localhost friction.

A CLI for managing coding-agent skills and MCP configs across tools. It gives you one place to inventory, audit, trust, and sync agent setup across Codex, Cursor, and Claude.

Built and shipped for a large live entertainment company. Real-time event data, custom agent behavior, and customer-facing product flows inside a larger system.

Internal AI systems for document-heavy workflows, sales intelligence, and RFP-related decision support inside a large industrial business. The goal was to reduce the amount of manual interpretation required to move from dense materials to usable operational signals.

A JSX-based templating system for generating prompts and other structured text with TypeScript. Useful when prompts need composition, reuse, and actual structure instead of giant strings.

Structured extraction for TypeScript LLM apps. It makes model outputs conform to schemas so the result is easier to trust, inspect, and use in real systems.

A streaming JSON parser that produces safe partial results against a schema while data is still arriving. Built for structured outputs that need to become usable before the stream is done.

A cross-provider LLM client that smooths over provider differences and keeps one interface across multiple model APIs.

Data and reporting systems for the Mr. Holland's Opus Foundation, built to turn scattered educational inputs into clearer reporting, artifacts, and internal insight. The work sat between analytics, workflow tooling, and AI-assisted report generation for a team that needed better visibility into its programs.

Built with Reshma and Moms First to make paid leave benefits easier to navigate across a messy patchwork of policies and government systems. It combined multilingual guidance, policy retrieval, and a clearer path through a process that is usually fragmented and difficult to understand.

Structured streaming responses defined with Zod schemas, so partial model output can be validated and used as it arrives.

React hooks for working with streaming responses, especially the awkward parts of LLM output in UI.

An open source LLM proxy and observability layer for seeing what your models are actually doing in production.

A TypeScript evals toolkit for scoring model behavior without leaving the rest of your application stack behind.

An MVP for private equity diligence work: structured intake, document-heavy review, and faster ways to surface risk, gaps, and follow-up questions across messy deal materials.

Front-desk style agents for dental scheduling and office workflows, designed to work across older practice systems that were never built for this kind of automation. The project focused on reducing coordination overhead while still fitting the constraints of legacy software and real office operations.

Real-time sports analysis and narrative systems built around live game data, shifting context, and the moments where information needs to become legible fast. The work sat somewhere between analytics, editorial tooling, and fan-facing product surfaces.

A small product experiment around membership, ordering, and lightweight operational flows. It was less about polished marketing software and more about making a niche workflow feel simple, legible, and fun to use.