Daniel Romitelli

I take whole systems, and the evidence is that they are still up. AutoLensAI walks a dealer around a car with an iPhone and renders a true interactive 360 from the photographs. Scenematic compiles a scene graph into generated video. NeuroLoq models the forgetting curve. Pramaana runs conformal prediction over 36 million rows of market data. This site is one of them too. Every one of those is mine from the schema to the pixels, and every one is live right now.

I also build inside other people's platforms. At MarutAI I was primary author of MAIDE, the developer environment they now sell, taken from an empty repo to a signed installer in four months. Three of the things it ships on are mine: Managed Changes, where an AI edit is proposed, approved by a human and kept as a durable audit record; the evaluation harness behind its release gates, which issues tamper evident signed certificates; and ManifestGraph, the lineage graph its agents query under fail closed semantics. I held one of two codebase wide approval seats, which meant reviewing work in parts of the system I had not written, and being one of the two people who had to say no.

Before that I was sole architect of an enterprise AI platform on Azure: 444 endpoints, 69 cloud resources, six autonomous agents and the framework that orchestrates them, held at 99.95 percent uptime with a 90 percent cut in inference cost once prompt caching went in front of it. There was nobody to hand the hard part to.

Four years operating P-3C acoustic sensor systems, tracking submarines from a patrol aircraft. The work was listening to noise and deciding what in it was real, on a schedule, with people acting on the answer. None of it was reversible at altitude.

That is why I test the failure path before the happy one, and why a system that overstates its confidence bothers me more than one that is slow. Most of what I write about now comes out of the same habit: making a system earn a claim before it is allowed to make it.

I build furniture and home additions by hand, from framing to joinery to finish. The gallery runs to more than 500 photos of real work: custom furniture, entryway and interior doors, kitchen cabinetry and bay garden windows, workbenches, shelving, trim carpentry, and a full basement buildout.

Most pieces start as a SketchUp Pro model and get rendered in Enscape, so a client can walk through the finished thing in photorealistic 3D before a board is cut. Parts that have to be identical go out to CNC. Under SteadyStone I ran the same process at building scale: remodels, large additions, multi-unit builds, and custom homes.

See the gallery →

Contract and consulting engagements where the whole system is on the table: architecture through release, with the evidence that it ran. I have proven I can carry a platform alone, and I do my best work where someone pushes back on a design before it reaches production. Cloudelipute LLC is where that work lives.

This site is the argument rather than a slide deck. Every demo runs live and every API call is real.