Doug McKay
Full-stack engineer and creative practitioner. I build the production systems that automate the Adobe stack.
Currently: orchestration on my own Hermes Kanban layer rather than n8n or ComfyUI · Qwen 3.8 on local hardware, which on my own workloads is now outrunning Gemini 3.1 · Fable 5 and Kimi K3 as the daily workhorses.
Adobe tools
Self-assessed, 1–10, master at the top. The second row is the rest of what I reach for — listed at what it honestly is, including where it is thin.
How the engagements run
The seam is the whole point. Plenty of people can build an agentic pipeline; plenty of others can stand in front of a marketing department. The engagements that stick need the same person to do both — because the flow you design is only as good as your understanding of the job it replaces, and that understanding comes from being the one asking the questions.
Current work — Firefly as the product photography pipeline
Premio Inc
premio-inc — AI Edge ComputersPurpose-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.
- Ingested their entire existing image library off their live site, then ran batch background removal and upscale through Firefly and handed the cleaned set back to the web design agent that builds the pages
- That is the whole point: the asset pipeline and the site generator are one system, so a catalogue migration is a batch job rather than a design project
- Google Veo for video generation alongside the Firefly stills
- Astro, deployed to Cloudflare Workers via Wrangler — static-fast at the edge, which matters when the pages are this image-heavy
- Third manufacturer on the same platform pattern — the reusable half is the point, not the individual site
- Cutout treatment, shadow and surface direction that holds across a catalogue of visually dissimilar hardware
- Next: extending into their marketing function for social content generation
- Then: building generation directly into the publishing pipeline, so campaign assets are produced as part of shipping a product page rather than as a separate request to a design queue
Same pattern, two more manufacturers
Teguar Computers
teguar.computerPurpose-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.
- Full commercial storefront on Astro / Cloudflare Workers — catalogue, product detail, configurator, quote flow
- Firefly background removal and scene generation wired into the asset pipeline rather than run by hand per image
- Art direction of every generated environment: lighting, perspective, scale and reflection have to match the product plate or a clinician spots it instantly
- Firefly is the default because the client already pays for Adobe. A tool the marketing team already has a licence and a login for is a tool they can keep using after I leave — that argument closes faster than any quality comparison
- Agentic training for their marketing and sales departments — a paid engagement teaching staff to hand their everyday tasks to agents
- Training on the custom systems I had already deployed for them, so the tools stopped depending on me being available
- The build created the credibility; the enablement is what made it stick
Cybernet Manufacturing
cybermed-g24 (in build)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.
- Catalogue, product detail and live configurator (OS, RAM, storage) with a quote flow
- One design system serving a second brand, so a new manufacturer is a configuration rather than a rebuild
- Specification data modelled so a SKU's certifications and I/O drive the page instead of being retyped into it
- A believability bar higher than consumer marketing — a clinician recognises a wrong ward instantly, and a wrong scene costs the sale
- Brand system distinct from Teguar's while sharing the same underlying structure
Cybernet Lead Intelligence
GIS lead-enrichment hub wired into an Adobe ColdFusion CRMA central intelligence hub for a manufacturer's marketing and sales function. An inbound lead arrives from the website; the system finds lookalike companies, sources and enriches contacts, scores buying potential, plots the whole target set on a GIS map, and drafts the actual approach — contact strategy, email sequence, call script — per account.
The part I am proudest of is the least glamorous. This did not replace anything. It was integrated into the company's existing Adobe ColdFusion custom CRM over REST endpoints and webhooks, so an enriched account appears inside the system the sales team already uses, already trusts, and already has years of history in. Nobody had to adopt a second tool.
- Lead ingestion from the website, lookalike-company discovery, contact sourcing and enrichment, buying-potential scoring across twelve verticals
- Apify network integration for data a marketing department cannot otherwise reach — the difference between a list and a pipeline
- Two-way integration with the incumbent Adobe ColdFusion CRM via REST APIs and webhooks, including de-duplication against records already in the CRM
- GIS map interface over the full target set, with filtering by vertical, geography and whether an account has been CRM-checked
- Generated per-account playbooks — contact strategy, email sequence, call script
- Integration beat replacement. A sales team will not move to a new system to chase a lead; they will act on one that shows up where they already work
- De-duplication against the existing CRM was the trust feature — a tool that surfaces accounts you already own gets switched off in a week
- The map is not decoration: territory is how a field sales force actually thinks, so the interface matches the mental model rather than the data model
Flagship — Chirps: an almost 100% Adobe render pipeline
Chirps
chirps-navy.vercel.appA 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.
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.
- Python orchestration driving After Effects, Premiere Pro and Media Encoder as scriptable render workers
- Agentic workflow triggered on user submission — no human touches an asset between upload and published video
- Web app for script authoring, avatar selection and scheduling
- Delivery tier went through three generations — Python driving Adobe Express, then a Premiere render server, then a headless ffmpeg + Python encoder as per-clip volume grew
- Originally served from Adobe ColdFusion; migrated to Astro on Cloudflare Workers
- Designed the proprietary on-screen video format — a recognisable, high-impact treatment styled after broadcast lower-thirds
- Compositing and keying decisions built into the AE templates the pipeline populates
- Avatar direction, framing and brand identity
adobe-render-bridge — an
aerender wrapper, ExtendScript generation for After Effects / Premiere Pro /
Media Encoder, a Media Encoder watch-folder driver, and a dry-run mode so the whole pipeline
runs on a machine with no Adobe software installed. It is browsable below.Browse the code
Production pipelines and internal tooling
USB-stick product photography pipeline
Built for an e-commerce companyA product photographer plugs a USB stick into a Raspberry Pi sitting on the desk. That is the entire user interface. The pipeline ingests every image on the stick, cuts out the product, removes dust and sensor spots, corrects lighting, and emails back a single link to the finished set. No login, no upload dialog, no training session, no new tab.
- Raspberry Pi as an ingestion appliance — mount detection triggers the run, so the physical act of plugging in is the submit button
- Batch orchestration with per-image retry, and a delivery link generated on completion
- Model-agnostic generation layer — built against the Gemini API ("nano banana"), and the call site is one adapter
- The team had already rejected two web tools. The blocker was never capability — it was that a photographer mid-shoot will not stop to use a web app
- Removing the interface entirely was the adoption strategy, not a shortcut
- Consistent output beats best-case output: every image gets the same treatment, so the catalogue looks like one catalogue
Adobe in production, elsewhere
DMARC Director
tangent.com/dmarcA production platform that audits email servers and enforces DMARC, SPF and DKIM, with hosted MTA-STS, BIMI, SPF flattening and TLS reporting. The Adobe angle is the go-to-market: Express and Firefly run inside agentic routines that produce the marketing team's social content, so campaign creative for a security product is generated on a schedule rather than commissioned.
YouPost
Internal social media automationSocial publishing automation built on Adobe APIs: Firefly image generation and video generation calls, Express for product background removal, and scheduling on top. Part of a broader pattern — I have built a number of internal marketing automation dashboards that sit directly on Adobe services and applications rather than around them.
Custom models, compliance and regulated domains
On-premise inference for regulated clients
Hardware, networking and compliance designSome 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.
- Low-cost, low-wattage GPU inference on Intel hardware — or Apple silicon — rather than a rack of datacentre accelerators the client cannot power or cool
- Running Qwen3.8 at 50–100 tokens/sec, which is hosted-frontier-model speed for the orchestration and execution workloads these flows actually do
- Speed stopped being the interesting number a while ago. On my own workloads that local Qwen 3.8 is now outrunning Gemini 3.1 — which is what makes the compliance answer viable rather than a sacrifice the client has to accept
- Custom management dashboard driving the on-site rack directly: model loading and placement across GPUs, thermal and utilisation telemetry, and live power draw
- Because power is the real objection, the dashboard also costs it — electrical spend hourly through yearly, and the solar array and battery capacity that would carry the load overnight
- The compliance answer to “where does our data go” is a room, not a paragraph in a vendor's terms
- Once inference is in-house, the cost model inverts — capital and watts instead of per-token billing, which is why the energy maths sits on the same screen as the throughput
- Same discipline as the audit work: I have done the cloud-side HIPAA and FedRAMP design too, so the boundary is drawn deliberately rather than by accident
CUBES AI
tangent.com/cubesA centralised license and subscription management platform — real-time tracking, renewal alerts up to 90 days out, unused and underused license detection, and audit-readiness. Built on a custom model trained on license data and compliance rules, not a general-purpose LLM with a prompt. No Adobe technology in this one; included because it is the clearest example of me training a domain model against a compliance corpus and putting it in front of paying users.
ScribesAI — original build
Healthcare dictation to chartsI helped build the technology behind an always-listening clinical dictation system that transcribed healthcare information into patient charts — LangChain orchestration over pre-trained open weights with fine-tuning on domain data. Healthcare and life sciences, with the data-handling constraints that come with it.
Gunhawk
Custom model, custom hardware, heavy complianceA project that required a purpose-built AI model, custom hardware, and substantial regulatory and compliance research before anything could ship. I built AI agents to carry each leg of the work — research, model, hardware integration, compliance — and drove it through to completion and sale.
Prism
matrix1.base44.appBuilt as a live tool for a California congressional candidate: aggregates data from many sources to model political outcomes and let a campaign target voters directly. The client has since asked to turn it into a SaaS product.
Azure Security Scanner (Scout)
Buy-vs-build, argued with numbersEnterprise-grade cloud security auditing at a fraction of market cost by wrapping ScoutSuite in a purpose-built dashboard and API. The interesting part was not the code — it was making the commercial case: Wiz, Orca and Prisma Cloud start around $20–30K a year for roughly 100 workloads and scale into six figures, against an open-source engine that produces the same foundational posture data. I owned the argument and then the architecture split across a three-person delivery team.