Matt
Maas

Does a compiled agent-context kernel preserve decisions at ~29% of the prose tokens?

Preregistered LLM context-compression benchmark. The primary finding replicated across 6/6 model×pack cells: a compiled agent-context kernel preserved decisions at ~29% of the prose-token cost. The secondary edge over naive truncation was model- and pack-dependent — reported honestly as a partial verdict rather than a win.

Answers real phone calls, stays grounded in what it can verify, and unlocks sensitive actions only for a verified owner.

Answers live Twilio calls with low-latency speech-to-speech AI - first on OpenAI/Azure Realtime, now AWS Nova Sonic over a Pipecat pipeline - and grounds its replies in a curated knowledge base instead of improvising. One codebase serves owner, guest, and legal personas; a PIN keyed in mid-call elevates the owner, and sensitive tools stay locked until it is verified.

The AI-assisted engineering workspace that the rest of this work is built in.

A working lab for production-quality AI-assisted engineering: task management, agent session memory, vector search over a personal corpus, and workflow automation. Legal, financial, and human-approval actions route through explicit safety gates instead of being left to the model.

Tap-chord input, AR-glasses display control, and a phone client for a self-hosted coding agent.

Native Android work built around hardware I actually use: TapLab replaces the Tap Strap 2's stock typing with a fixed-length two-chord alphabet and a data-driven trainer, and an XREAL toolkit turns a rooted handheld with AR glasses into a pocket desktop. It also ships a ~250-line Kotlin client that drives a self-hosted AiderDesk instance from a phone. 3 APKs published via GitHub Releases.

Cross-platform legislative intelligence app with AI-generated bill summaries.

Cross-platform mobile app that makes U.S. Congressional bills readable: browse and search legislation, with AI-generated summaries and key-change callouts drawn from Congress.gov data. A .NET MAUI client over an Azure pipeline - Cosmos DB storage, Azure Functions orchestration, and OpenAI summarization.

Local-first email triage that keeps the mailbox, and the rules, on your own machine.

Local-first Gmail triage and automation: sync a mailbox into a local cache, then classify, label, and organise it with LLM assistance. Multi-provider model routing, including local models, keeps the pipeline from depending on a single vendor.