I build systems for places where failure isn’t recoverable.

Bloom was a vertical-farming stack built around one question: can a machine see disease earlier than a person walking the aisles?

We trained convolutional image models on leaf imagery to classify disease and deficiency from the visual signal alone, then fed the classification straight back into the hydroponic control loop. Nutrient dosing, EC and pH targets, photoperiod and spectrum all moved on the model, so the response was automatic rather than advisory.

This was 2016. There was no off-the-shelf vision API worth the name for this, no managed training platform that made sense at our budget, and no pretrained agricultural model to fine-tune. We gathered the dataset, labelled it, trained it, and ran inference on hardware sitting in a humid room next to the plants.

OneIsland compressed the distance between designing an off-grid dwelling and manufacturing one.

The core was a web-based parametric CAD environment: geometry kernel, constraint solving and real-time visualisation running in the browser, coupled directly to CAM output and robotic fabrication. A change to the design propagated through to toolpaths and cut files without anyone redrawing anything.

The hard part was never the rendering. It was building a geometry system that stayed numerically stable across thousands of parametric edits, that emitted toolpaths a machine would actually accept, and that treated the physical constraints of living off-grid as first-class parameters rather than notes in a PDF: thermal envelope, water capture, power budget, structure.

DiveLogic is decompression planning for technical, cave and closed-circuit divers. It is the segment where the plan is the safety equipment.

It implements Bühlmann ZHL-16C with gradient factors and VPM-B, tracks CNS, OTU and IBCD across every phase of a dive, and plans multi-gas trimix and heliox profiles with automatic switch depths, bailout ascents, lost-gas contingencies and rule-of-thirds reserves. It syncs over Bluetooth with Shearwater Perdix, Petrel and Teric computers, and replays tissue loading afterwards so a team can debrief against what actually happened.

The AI sits on top of the physiology, never in place of it. Divers interrogate a plan in natural language and the system proposes adjustments, but every number it returns is produced by the deterministic model underneath and traceable back to it. In this domain a plausible answer is worse than no answer.