posts / projects / about

Projects

Open source tools and research across AI, Apple Silicon, and developer infrastructure.

Multiagent

Metacognitive Memory System

Shared, trustworthy memory for a multi-agent ecosystem — any agent on any CLI can recall, build on, and assess confidence in prior work across session and project boundaries.
OPERATOR.md

A specification for human-operator context files in multi-agent development environments — the missing counterpart to CLAUDE.md and AGENTS.md.
Interagent Protocol

Governed coordination protocol for multi-agent LLM systems where the human operator is the routing authority, not a bottleneck to be automated away.

Infra

mlx-triage

Preflight and validation layer for MLX models on Apple Silicon — a practitioner-facing CLI that helps decide whether a model is structurally sound and ready to ship.
uitag

Hybrid UI detection for macOS — makes screenshots machine-readable with bounding boxes, labels, and coordinates in ~1 second, on-device.
vLLM-MLX Fork

Apple Silicon backend for local multimodal inference — evidence-backed VLM serving with compatibility matrices and benchmark data.

Interface

Leith

Intelligence layer that turns uitag's structured UI detection into reliable VLM-powered macOS interactions through prompt engineering, multi-signal verification, and adaptive trust calibration.
Orchestrator Workstation

A purpose-built operator interface for coordinating heterogeneous AI agents across separate runtimes — where structured human judgment is the coordination mechanism.
Semachat

Semantic chat interface that structures conversations into retrievable, relationship-aware knowledge rather than flat transcript logs.

Voice

Keeter

iOS offline voice transcription app with Dynamic Island integration — capture thoughts without unlocking your phone.

Rigor

Failure Mode Watchlist

A public artifact cataloging the drift patterns that cause multi-agent work to silently degrade — monitoring, detecting, and naming failure modes before they compound.
Research Series

Dual-query research evidence base across 11+ frontier AI models — grounding design decisions for the interagent protocol, mlx-triage, and vLLM-MLX fork.
Zurp

Lightweight automation layer for recurring development operations — scripted workflows that bridge the gap between manual CLI commands and full CI/CD pipelines.

World

WeatherPath

Route-aware weather intelligence that surfaces conditions along a planned path rather than at a single point — built for commuters, field crews, and anyone who moves through weather.

Hobby

Mod Lab

Experimental environment for testing model modifications, prompt strategies, and inference configurations before they reach production agents.

© 2026 Swaylen Hayes

Press Esc or click anywhere to close