Open navigation
Work / Selected systems
Selected systems.
Architecture, infrastructure, failure modes, deployment, and operating details for each system.
Production AI systems
Building nowAI Software Factory
A deterministic software factory with sandboxed workers, a deny-by-default tool gateway, and an idempotent publisher outside the model loop.
The model plans. The control plane decides.

Voice AI
Mossaic
Voice-first research recall. Ask the corpus out loud and receive cited answers in under a second, backed by one shared index across voice and Slack.
Sub-10 ms retrieval with source citations.

Systems engineering
Argus
A terminal control plane for agent sandboxing written in Rust. Every dependency had to earn its place; every state transition is visible.
3.7 MB binary. Nine dependencies. No async runtime.

Agent orchestration
Silicon Cortex
Event-driven AI workers on Kubernetes. Ephemeral pods receive work, generate code, publish results, and disappear when the queue is empty.
Scales to zero. Each run emits an artifact and provenance record.
Developer tools
ArchByte
Real-time architecture maps for any codebase, assembled by parallel agents that inspect code, infrastructure, and documentation.
Maps code, infrastructure, and documentation in one graph.
Maps code, infrastructure, and documentation in one graph.
Applied AI
localhost
A real-world scavenger hunt against alien agents that turns the player’s neighbourhood into the game board.
GeoGuessr meets Pokémon GO, played from the couch.

Distributed systems
Meridian
A RoboTaxi simulation where pricing, dispatch, location, and language-model services coordinate over an event stream.
Ten services coordinated through Kafka.

The wider archive