engineering journal

BuildLogs

Short, honest entries from the bench. Every build breaks something, and this is where I write down what and why.

Jarvis2026-06-18

ChangedAdded hand-tracking and object-recognition experiments to the Jarvis pipeline.

BrokeLocal laptop performance became a bottleneck, running vision and a local model together choked the frame rate.

LearnedOptimization matters when running vision and local AI together; you can't just stack two hungry pipelines and hope.

NextImprove model and pipeline efficiency before adding gesture control.

Jarvis2026-06-09

ChangedGave the quant and vision subagents their own sub-orbs that orbit the core and change color by state.

BrokeState updates were re-rendering the canvas and stuttering the animation.

LearnedRead state once per animation frame instead of through React, the render loop should never depend on the component tree.

NextGeneralize the sub-orb system so any subagent can register one.

Jarvis2026-05-28

ChangedSwitched the orb to additive 'lighter' compositing on pure black so unlit pixels read as 'no light' on the wall.

BrokeCalibration crosshair geometry drifted whenever the window resized.

LearnedOn a projector, black is the absence of projection, so design for light you add, not pixels you draw.

NextAdd a rule-of-thirds calibration grid for wall alignment.

Automation2026-05-16

ChangedWired Jarvis to room lighting so it can set scenes and react to events.

BrokeRate limits, rapid scene changes got throttled by the device API.

LearnedDebounce intent, not just requests; the user means one outcome, not ten.

NextMap more of the room into a single addressable environment.

Biomedical2026-05-04

ChangedPrototyped a rules-and-evidence engine that turns genomic markers into traceable insights.

BrokeEdge cases in conflicting markers produced contradictory outputs with no resolution order.

LearnedDeterminism needs an explicit precedence model, 'it depends' has to be encoded, not implied.

NextAdd an audit trail so every conclusion links back to its evidence.

CAD2026-04-22

ChangedModeled and printed the first version of the laptop rig mount in Fusion 360.

BrokeFirst print didn't fit, clearances were 0.4mm too tight.

LearnedMeasure twice, then add tolerance for the printer's reality. CAD lies about fit.

NextAdd ribs for stiffness and reprint at corrected clearance.