Intent
The message is matched to the agent and skill that own it.
Selected work
Useful software with a point of view.
I build AI systems, Mac utilities, civic tools, and native apps around one habit: turning loose context into interfaces people can actually use.
- 26
- apps
- 26
- AI covers
- 44
- captured screens
AI-native products
Local Mac utilities
Native app experiments
Systems with receipts
Selected portfolio
A sharp edit of the apps worth showing.
Each card opens the cover art, captured screens where available, and the build context behind the product.
The system behind the work
JARVIS — the operating system that builds the rest.
Every app on this page runs through one orchestrator. JARVIS reads the request, routes it to the specialist that owns it, works against layered memory, and leaves a receipt behind. It self-configures on first run and rewrites its own playbook as it learns.
- 60
- specialist agents
- 11
- prime directives
- L0–L5
- memory stack
- 210+
- engineering skills
- ~350
- token cold start
- 24/7
- always on
Route
The orchestrator delegates with clear instructions. It coordinates; it doesn’t execute blindly.
Build
The specialist runs with the most direct tool and relevant past lessons injected first.
Receipt
Cortex records the handoff, shared context, and proof so work survives across sessions.
Remember
The memory loop persists what changed and what was learned for next time.
Route, don’t execute
Intent maps to the agent that owns it. JARVIS delegates and synthesizes instead of doing everything in one overloaded thread.
L0 identity to L5 raw
A six-layer stack that loads lazily — a ~350-token cold start, with semantic recall pulling the right files only when the topic calls for them.
Coordination with receipts
Durable queues, shared context, and accountable handoffs across sessions, plus a scout that goes and finds a missing capability.
A library that compounds
210+ engineering skills and a growing set of learned lessons. A pattern seen three times gets promoted into a reusable skill.
The most direct tool
Direct connectors for mail, calendar, docs, chat, and the web, plus native control of the Mac — always preferred over slower automation.
Learns from its own misses
Failed tool calls are distilled into lessons automatically, and state is saved before the context window ever compresses.
The team it routes to
One chief of staff over a roster of specialists, each tuned to a model and a domain.
Orchestrator
Researcher
Content Strategist
Scheduler
Finance
Builder
App Studio
Analyst
Web Designer
Onboarder
Independent Grader
+ ECC builder sub-team
How it operates
Memory is sacred
Read at the start of a session, written before the end. Nothing learned is allowed to evaporate.
Plan before acting
For anything substantial: propose, get explicit sign-off, then build. No build without the go-ahead.
Minimal scope
Do exactly what was asked. Don’t touch unrelated code, and don’t improve what wasn’t broken.
Protect context
Summarize large outputs, clear between phases, and send sub-agents to explore so the main thread stays sharp.
Boring is beautiful
A deterministic script beats an agent nine times in ten. Reach for judgment only when the work truly needs it.
Discover before blocking
When a capability is missing, go find it — a tool, a skill, an MCP — before ever saying it can’t be done.
Operating style
Find the messy moment, then design the handle.
Across the work, the pattern is practical: compress scattered inputs into a clear surface, give the user a next move, and leave enough proof that the system can be trusted.