Matt Rickard's Systems Approach to AI Tools

Table of content

Matt Rickard doesn’t just use AI—he builds the plumbing that makes it actually useful.

As a former Google engineer who worked on minikube, skaffold, and Kubeflow, Matt spent years making distributed systems accessible to developers. Now he’s applying that same systems thinking to AI, creating open-source tools that solve real problems instead of chasing hype.

The Problem with Raw LLM Output

Anyone who’s tried to integrate LLM responses into production code knows the pain: models hallucinate, return malformed JSON, ignore your format instructions. Matt’s answer? Don’t ask nicely—constrain mathematically.

ReLLM forces LLM output to match a regex pattern by manipulating logits during generation. Want a date in mm/dd/yyyy format? The model literally can’t produce anything else:

Pattern: [0-9]{2}/[0-9]{2}/[0-9]{4}

It sounds simple, but the impact is huge. Even tiny models (GPT-2, 124M parameters) become reliable when they can’t deviate from the expected structure.

ParserLLM extends this to context-free grammars—perfect JSON, valid XML, whatever structure your pipeline needs.

Browser as the New Runtime

Matt sees browsers as the logical place for AI inference. Privacy benefits, no API costs, offline capability—the arguments are compelling.

@react-llm packages the complexity into simple React hooks:

const { send, conversation, init } = useLLM()

Behind the scenes: WebWorkers handle off-thread inference, models download from HuggingFace CDN, WebGPU accelerates computation. The live demo at chat.matt-rickard.com runs Vicuna 13B entirely client-side.

Unix Philosophy for AI

Matt wrote an influential piece translating Unix philosophy for the AI era. The core insight: LLMs should be composable, like pipes.

His modified rules:

This isn’t abstract theory. His openlm library lets you call any LLM with OpenAI-compatible syntax. llm.ts does the same for TypeScript across browser, Node, and Deno.

Personal AI Workflows

Matt documents how he actually uses LLMs in daily life:

Grocery list organization. He appends items throughout the week, then uses a prompt to sort by store section. Simple but saves real time.

Book note compression. Highlights from Apple Books get extracted via SQLite query, then summarized by an LLM. The workflow chains existing tools—very Unix.

Dictionary lookup. Unknown words collected while reading get batch-processed into formatted definitions.

Task decomposition. Not for creative tasks—LLMs aren’t great at those—but for breaking known tasks into subtasks.

Writing editor. Critical feedback mode, not generation. He finds critique more useful than raw text.

He’s transparent about what doesn’t work: project naming (too clichéd), brainstorming (can’t escape obvious paths), fiction writing (still experimenting).

Daily Writing Practice

Matt has blogged daily for over 700 days. Topics range from technical deep-dives to business strategy to classical references (Cincinnatus, Schopenhauer’s argumentation strategies, the Catilinarian Conspiracy).

Posts worth reading:

Standard Input

His current project builds AI-enhanced personal productivity apps:

The pattern: take boring personal tools, add thoughtful AI features, ship on web and iOS.

What Makes His Approach Different

Matt treats AI like any other engineering problem: understand constraints, build tools that compose, document what works and what doesn’t.

No grand proclamations about AGI. No breathless product launches. Just systematic building, daily writing, and sharing what he learns.

Key resources:

His system instructions page explicitly welcomes automated systems—peak engineer mentality.