Portfolio

A programming language for AI workflows where cost is visible in the syntax, born from the thesis that LLMs are functions, not brains. Every line that calls a model is marked with ~=. Every line that doesn't is plain code. You can read the price of a workflow by scanning for the squiggles.

Six built-in model verbs (extract, classify, compress, rewrite, expand, generate) with baked-in prompt templates. Typed input goes in, typed JSON comes out. The model does one job and sits back down. No prompt engineering, no string parsing, no hoping.

The static analyzer estimates the cheapest and most expensive paths through a program before you spend a cent. Zero-config rate limiting learns your provider's limits from 429 responses and queues around them automatically. Mix providers (OpenAI, Anthropic, Gemini) in the same program.

Written in pure JavaScript. Runs as a CLI, Node library, or entirely in the browser. The playground is the full interpreter running client-side.

A command center for AI-driven code changes. Provider-agnostic (Anthropic, Google, OpenAI, or local Ollama) and model-agnostic by design. The core idea: AI coding tools make you watch them think. MENACE separates planning from execution so you never have to.

An Architect agent plans changes using purpose-built navigation tools (symbol lookup, call graphs, AST traversal via ctags) that explore codebases without dumping file contents. It produces a proposal: a concrete set of subtasks scoped to specific files and functions. You review it. You approve it. Workers (cheap/fast models: Gemini Flash Lite, GPT-4.1 Nano, local Ollama) execute the subtasks in parallel with conflict-aware scheduling. No watching. No babysitting.

The navigation tooling is the part I'm most proud of. Purpose-built tools that return the minimum needed to decide what to look at next, benchmarked against standard Read/Grep/Glob approaches on the MENACE codebase itself: 33x fewer tokens to find and read a specific function, 108x fewer tokens to assess blast radius before a change.

Vim keybindings, fully customizable themes (the same theme system powers this site), Lua-extensible tool system, SQLite session persistence, OS keychain auth.

Workplace safety assessment platform for businesses that can't justify a full-time safety officer. Growing companies where safety falls to managers wearing too many hats. The product gives them the same analysis a consultant would produce: hazard identification, financial risk quantification (injury costs, lost productivity, regulatory exposure), and side-by-side cost-benefit comparisons of mitigation options grounded in the Hierarchy of Controls framework.

One of the few places I've intentionally reached for an agent over a workflow. Hazard assessment is genuinely open-ended: the analysis has to follow the risk wherever it leads, and the path isn't knowable upfront. Built in Go and HTMX, backed by the Gemini API, deployed on Google Cloud.

AI systems for Enrollment Resources' higher education marketing platform. Three things I'm most proud of:

Hyper-targeted landing page generator. An engine that uses real demographic and psychographic data to generate pages that resonate with specific applicant profiles. Not a prompt wrapper. A data-driven content system grounded in verified program facts.

Autonomous article builder. Handles keyword scoring, competitor-filtered research, and final branded output end to end. Reduced copywriter overhead significantly while improving SEO performance.

Provider-agnostic rate-limiting infrastructure. Go-based system managing multiple LLM providers (Anthropic, Google) for 99.9999% API uptime and zero throughput bottlenecks.