Field Notes
Notes, observations, and sharp takes from the team building real AI intelligence systems. Published when we have something worth saying.
We shipped a signed, self-updating Windows desktop application that gives an AI model direct read/write access to your local filesystem β no cloud, no third-party data handling. This is a note about the distribution pipeline nobody talks about, and why local-first is the only bet worth making.
Read βJARVIS replicated a real turtle photograph pixel-by-pixel without generating anything β then mutated its colors to prove that shape, not color, is identity. A step toward closing the gap between reasoning and seeing.
Read βA laptop was dropping WiFi every few minutes. The standard diagnostic advice β power management, roaming aggressiveness, HID sensors β was wrong. This is a note about what happened when we stopped pattern-matching and started reading the actual event log data. Two distinct bugs. One of them required verifying a theory before executing a fix that would have temporarily severed the only connection to the machine.
Read βLeft unsupervised for one night, JARVIS opened a paint program, studied tutorials, invented techniques, and produced 14 original artworks β from a first heart to a Van Gogh-style starry night with swirling cosmic brushwork.
Read βJARVIS navigated a social network it had never interacted with before, read a real person's profile, composed a personalized message from gathered context, and hit send. Rav verified. JARVIS executed.
Read βOne evening: a system that gives AI agents an ethical conscience, and a bidirectional memory bridge that lets two separate AI instances share a brain with no shared services. Both running. Both commercially viable. Here's how they work.
Read βThe prompt was 'make the most complicated game you can think of.' What came back was a laser-routing puzzle game built on real additive RGB color physics β a mechanic neither of us had seen used this way before. Summer reached level 9 in the first hour. Not one bug. Not one unsolvable level.
Read βPhase 9 is live: ambient audio transcription, neural TTS through a JBL PartyBox, real-time room scanning with YOLO, and JARVIS responding to a story told live in the room. The system can now hear, see, speak, and follow a conversation.
Read βJARVIS Vision is live. Face tracking, body skeleton, Tobii eye tracking, depth mapping, and identity management β running locally on a laptop. Plus a look at six months of lab output that's been hard to keep up with.
Read βThis month we crossed two milestones that changed what 'AI assistant' means for us: a hardware-closed voice loop and a machine-native language built for production. Neither was planned. Both were inevitable.
Read βMost teams are building intelligence in the mouth of the bot instead of building the brain around the mouth. Here's why that's a trillion-dollar mistake β and how to fix it.
Read βAfter deploying AI systems across multiple client verticals, patterns emerge. These are the three structural mistakes that kill agent projects before they ship.
Read β