Spend
6.69 Mb/s
One frame from my encoder comparison, encoded at CRF 20. PSNR 44.22 dB against the source.
Return
5.30 Mb/s
The same frame, one setting changed: 21% fewer bits, PSNR 43.01 dB. Can you see it? The highest-payoff moves usually look like this: small, cheap, and first.
Open the report
The first move pays the most.
So the judgment is the order: what first, what next, what to ignore.
Tools
5 tools I built, 4 of them run in your browser. Real measurements, and AI doing the work it is good at: looking, scoring, drafting, ranking.
Interactive/tools/payoff-ranker
Which job should AI take first? List 3 to 5 jobs your business repeats: how often, how long each takes, what an hour costs and the share a model could take. The ranker turns them into hours and dollars a year and puts the first move on top. Your numbers only, blank to start, and nothing you type leaves the page.
AI/tools/ai-video-quality
How does this video look to a person? The old answers were a viewer panel (slow, expensive) or PSNR and VMAF (fast, often wrong about what people notice). This pipeline asks a vision language model instead: FFmpeg samples the frames, libvmaf scores VMAF, PSNR and SSIM in one pass, and the model rates each frame 1 to 5 with a written artifact check, blockiness, banding, blur, returned as JSON beside the metrics. The expensive, repeated, human job is the one to automate first.
Interactive/tools/encoder-comparison
One x264 setting moved 2 points, CRF 20 to 22: 21% fewer bits, PSNR 43 dB, and you cannot see it. Wipe between the frames, then open the per-frame PSNR series and the FFmpeg commands behind it. A first move in one picture: small, cheap, measured.
Interactive/tools/cdn-cost-model
Egress is the bill that grows when streaming works. Set subscribers, viewing hours and the bitrate ladder, pick a scenario (HEVC or AV1, per-title encoding, multi-pass transcoding) and watch the monthly number move. The shape every AI payoff estimate should have: your inputs, visible assumptions, a figure you can defend.
Interactive
Proposals and RFP answers assembled from approved text blocks and templates, exported to Google Docs or HTML. No model in this version: it automates the assembly and leaves review to a person. The natural upgrade is a model drafting the first pass from the approved library, with the same human sign-off. Quotes that take days sit near the top of most payoff lists.
Not publicly available
Work
Video and cloud work, and each piece was a decision an AI buyer faces now: which job to automate first, whether to buy or build, and when a pilot is ready for production.
Case 01
Architected cloud media workflows for major media companies: content processing, storage and distribution for broadcast and streaming, moved off fixed infrastructure in stages. What moves first and what waits is the same sequencing call an AI rollout has to make.
Case 02
Led 4K HDR broadcast and streaming projects for major operators, from lab evaluation to live service. A pilot that works is not a product that runs. The gate between the two is where AI projects stall, and this is what passing it looked like at scale.
Case 03
Designed and implemented automated processing pipelines on AWS Lambda, replacing manual ingest with event-driven workflows. A build-versus-buy decision in practice: a managed encoding service for the heavy lifting, custom orchestration around it. The pattern holds for AI today: buy the engine, build only the glue, and start with the queue a person is feeding by hand.
Publications
5 write-ups from 2019 to 2022, each linked to where it was published. Written for media teams, and the calls inside them are the ones an AI buyer makes now: automate the checks first, buy the platform, move in stages.
Explores how AWS partners Telestream and Ateliere integrate with AWS services to automate media ingest and ensure quality control in media workflows.
Highlights the use of AWS services in conjunction with Imagine Communications' SureFire platform to deliver broadcast-quality connected TV advertising.
Discusses the gradual transition of media workflows to the cloud, emphasizing the benefits of scalability and flexibility.
Describes the integration of Bitmovin's encoding solutions with AWS services to deliver high-dynamic range (HDR) video streaming for the HPA Tech Retreat 2021.
Provides a tutorial on integrating AWS Lambda with Bitmovin's Cloud Video Encoder to automate video ingest workflows, enhancing efficiency and scalability.
Independent
6 small products with public links. 4 of them run on AI somewhere: a model writing listings, reading requests, drafting code, or running a store's daily operations.
01
A patented magnetic cable tie, funded on Kickstarter at 700% of goal, sold on Shopify and Amazon. Since mid-2026 the store runs day to day on a fleet of 11 scheduled AI agents: SEO and answer-engine work, GA4 and Search Console pulls, Shopify edits with backups, TikTok drafts rendered in Remotion, and a second model that approves routine changes. I decide on money, public posts and customer data. My test bench for running a small business with agents.
02
Photograph everything you are selling and a vision model (GPT-4o-mini) writes each listing, what it is, material, condition, features, and proposes a price. Out comes a sale page with a cart and a chat line to your phone, plus bulk-upload CSVs for Facebook Marketplace, so the same 53 items go up in one batch instead of one form at a time. Enter your minutes per manual listing below and see what the batch saved.
Time saved, your numbers
Enter how long one listing takes you by hand.
03
Describe the look you want for any website and a model writes the theme, colours, type, spacing, as CSS you can edit live, with saved themes and a preview before you apply it. A small test of what a model does with a vague brief and a strict output format.
04
An Uber-style marketplace for pet waste removal: customer and scooper dashboards, Stripe payments, automatic onboarding and notifications, and a Telegram bot and post analyser that read requests with OpenAI through Supabase functions. The whole MVP was written with an AI coding tool (Cursor): a working test of how much software one person can ship now.
05
Year-by-year retirement projections that mix conventional savings with Bitcoin holdings and take-profit rules: an interactive chart, a full table, every assumption editable. A model with its assumptions in plain sight.
06
Checkout optimization for online stores: fewer steps between the cart and the receipt, measured on conversion.
I'm a technology operator based in Littleton, Colorado. I've spent 20 years designing, selling and running large-scale systems for global organizations, and I've built and run a few small products of my own.
AI pays in more places than most owners expect, so the question is rarely whether. I write about which job pays most in a business, what to do first, what comes next, and what to ignore. The questions are the same whether the business is a practice with a front desk or one division of a large company. I've also put AI into a small business and watched what it did to the numbers.
My background is cloud architecture and large-scale infrastructure, much of it in media and streaming, including senior global roles at AWS. Most of that work came down to the decisions AI buyers face now: which job to automate first, whether to buy or build, when a pilot is ready for production, and when to move on to the next one.