Doug McKay

Some clients cannot send their data to a hosted model. Under HIPAA or FedRAMP the agentic flow has to run inside their own boundary, which turns an AI project into a hardware, networking and audit problem before it is a software one. I do that whole side: specify and build the GPU inference hardware, design the network around it, and run the cloud HIPAA and FedRAMP compliance auditing and architecture for the parts that do stay in the cloud.

Actual build A 6× GPU on-site server on the bench, running a thermal soak before it ships. The Lenovo P8 compute nodes hold the dual Intel Arc Pro B70s you can see loaded in the fleet manager below — same machine, same cards, two views

Delivered as a unit, not a parts list

• The client's compliance boundary is a rack in their building — firewall, switching, GPU compute and storage specified, built and burned in together
• Thermals are the failure mode nobody catches on paper. Sustained inference load is not a spec-sheet workload, so every build gets soaked here before it leaves
• It arrives commissioned. Nobody on the client's side has to assemble an AI appliance out of purchase orders

Live rack telemetry — 487 W at the wall, and what that costs hourly through yearly, with the solar array and battery bank sized to carry it

Fleet view — four GPUs across three machines, local and over the LAN, with per-card model loading, VRAM, thermals, clocks and power

A marketing team will not adopt a tool that requires somebody to sit in Firefly all day. So the generation stays in Adobe, where the brand already has a license and a login — and everything around it is automated. I train a Firefly custom model on the client's own product photography, then my own harness, Hermes, runs it: work arrives as cards, an agent picks one up, drives the custom model, and the finished assets come back on the card.

The Adobe side — a trained custom model (Industrial PCs candidate) generating from a reference image, with the concept token CUSTOMSUBJECT carrying the client's product through every prompt

The automation side — agents holding two live Firefly jobs: generating product imagery against the custom model, and preparing the next training dataset. Nobody is clicking Generate

Purpose-built edge AI systems for machine vision, robotics and physical AI — NVIDIA Jetson, Intel Core Ultra, IP66-rated rugged platforms that put inference where the data is created. A deep SKU catalogue where every unit has to be photographed identically, and none of them are pretty out of the box.

Firefly does the product library. Every unit is upscaled and cut out of its original shot, then presented on a consistent surface across the whole catalogue — so a page of five edge computers reads as one product family rather than five different photographers' work on five different days.

Category hero — the unit cut out of its original plate and placed on a generated gradient surface

Product grid — five SKUs, one consistent cutout treatment across the catalogue

JCO-6000-ORN — upscaled cutout on a generated circuit-trace environment

“Look closer” — the same unit at full resolution, which is what the upscale step is for

Purpose-built computers for the places ordinary hardware does not survive — medical panel PCs certified to UL 60601-1, IP69K washdown industrial units, rugged edge systems. Every SKU needs to be seen in the environment it was built for, and photographing that is not viable: three shoots per unit, in facilities you cannot get a camera crew into. So each product is shot once, clean, and Firefly removes the background and generates the deployment scene.

Split-environment hero — two products composited into generated industrial and clinical scenes

NVIDIA edge AI — on-premise inference on the factory floor. 0 bytes sent to the cloud

Guides — every editorial illustration generated: an operating theatre, a washdown line, a rugged test lab, an IP-rating diagram, a fanless cooling cutaway. Firefly first, with nano banana where it fits

The same platform pattern, second manufacturer — and the clearest healthcare case. The CyberMed G24 is a fanless, antimicrobial, UL 60601-1 certified all-in-one for active hospital wards: sealed bezel so clinicians can clean it with aggressive sanitizing sprays, dielectric isolation on every port. The page has to put that unit in a ward, and the ward is generated.

CyberMed G24 — product detail and live configurator; the hero places the unit in a generated hospital ward

Edge AI — “AI that never leaves the building.” Dual RTX PRO 6000, 120B-class models, zero cloud egress

Stills are only half of it. For motion I pick the model inside Firefly and generate there, work the frame with Firefly's own edit tools, then take the clip into Adobe Express to cut it. The last pass is a person, every time.

Firefly — choosing the model, then working the generated frame. Upscale at 2× to 2048×1610, alongside Fill, Remove, and Crop and expand. Note the credit bottom right: powered by Photoshop

Adobe Express — the generated clip on a timeline at 1280×720, sliced to 8.6 seconds. This is where a person cuts around the frames where the model got the hardware subtly wrong

A platform that turns a written post into a hyper-realistic lipsynced video asset in a distinctive on-screen format, then schedules it across social networks. The product is the video. The engineering is the render pipeline behind it — and that pipeline is Adobe applications driven as headless infrastructure by Python.

Built under commission for a venture capitalist selling technology into news media and film studios. That shaped every decision in it: the output had to match a broadcast on-screen treatment rather than a social-native one, and it had to come out the far end at a volume and consistency a newsroom would accept — which is why the render chain is Adobe applications on rails instead of a person in an editor.

chirps-navy.vercel.app — the on-screen format the pipeline renders, and the avatar/copy treatment it populates per post

Orchestrated by Python. Every stage above except distribution is an Adobe application or service being called as a build step.

That progression is the point, not a retreat from the stack. The creative applications are where the treatment gets designed and where a human can still open the file and change it. The encode tier is a commodity, and once volume made it the bottleneck it belonged in a headless process — which is the same reasoning behind Media Encoder watch folders and behind Firefly Services existing as an API at all. Knowing which half of a pipeline is craft and which half is throughput is most of this job.

Output of the Express stage, in my Adobe library — source clips matted and replaced with a green screen so the After Effects template keys the same way every time. Two frames here are keyed; two still show the original plate with the matte edge drawn, which is the before/after of that step.

A RIP is where a designer's file stops being artwork and becomes ink on a substrate: color transforms, ICC profiles, halftoning, rasterization, and the device controls that decide whether the result matches the proof. BendiRaster is mine — an independent raster engine and color pipeline I built from the ground up, matching what Adobe PostScript does and then going past it on the axes above. BendiRip is the commercial product it ships inside, and it is live and in market, so the engine can be judged on running software rather than on my description of it.

BendiRip Studio driving an HP Latex 570 — job queue and cost per square foot, ICC profile and substrate/cure controls, live printer, ink and hardware telemetry

Two months solo is only possible if the work is graded. I do not point one model at a problem and hope. I decide what tier of thinking each layer actually needs, and route it there.

Substrate, ICC profile, pass count and cure temperature — the color decisions a print operator actually makes, surfaced in one panel instead of buried