Field Notes — NAB Sunday

Cloud Code, GitHub,
and a Lot of Late-Night Engineering

A walk through the website-creation conversation that started in the After Hours room on NAB Sunday — plus the other technical threads that filled the rest of the night.

Date Sun 2026-04-19 Window 09:10 – 23:00 CDT Source Room recap Setting NAB Las Vegas

By the time the convention floor closed, the After Hours room had picked up its second wind. The conversation kept circling back to one thing: how cheap, fast, and unreasonable it has gotten to build real working software in an afternoon — and what that means for everyone with a Google sheet full of gear, a YouTube channel, or a workflow that doesn’t quite exist yet.

Part One

Website & App Creation with Cloud Code and GitHub

Three different stories about building software the new way surfaced over the day. None of them were the same person, and that was the point: it’s no longer one wizard in the corner. It’s the room.

06:00 PM CDT · After Hours, NAB Sunday

“Just an app I put together with cloud code” — the Gear Ops demo

Asked what was on his second screen, an after-hours panelist shrugged and said it was “just an app I put together with cloud code — I call it Gear Ops.” He had fed it a Google-sheet dump of all his equipment, IP addresses included; the app listed everything by category, pinged each device to confirm it was online, and gave him a one-click browse button that opened the device’s admin page in a browser.

Why it landed: anyone in a control room knows the moment when a camera needs a tweak and the answer to “what was that IP again?” is a sticky note on the back of a monitor. Solving it once, in code, took an evening.

He was planning to show it off at the AI meetup on Friday night.

Claude CodeGoogle Sheets → JSONHTTP pingCompanion / ATEM / HyperDeck
09:25 AM CDT · Office Hours, morning session

The OHG back-of-house: an Amazon-affiliate gear database, vibe-coded

The morning’s funding conversation took a hard turn into engineering when one of the OHG developers walked the room through a months-long project: a panelist-pages system that turns each crew member’s gear list into a single, deduplicated affiliate catalogue.

The flow he described:

  • You type a product (e.g. Shure MV7+) and the system queries Amazon, preferring listings sold by Amazon directly or the manufacturer.
  • A second service (Serp API) returns the ASIN, image, and key metadata.
  • The catalog row gets a category (microphone / camera / accessory / rack), one canonical affiliate link, and a thumbnail.
  • Multiple panelists who own the same item all reference the same catalog row — one link, one place to maintain.
  • New items hit a pending queue; a human approves image + description + link before they fan out to every panelist’s page.

The build was Claude in the editor + Vercel for deploy + Supabase as the back end. He estimated the AI spend at a few dollars to generate hundreds of catalog pages. The repo started out private — “everything I put up has only ever been for me, so I always make it private” — and is being prepared for public so the community can extend it. That move triggered a careful key audit before the visibility flip (more on that in Part Two).

ClaudeVercelSupabaseSerp APIAmazon AffiliateGitHub (private → public)
06:25 PM CDT · Creator Lab stage, NAB

Vibe-coding plugins for software that doesn’t support them yet

Ed Dawson of CG Pro, presenting in the Creator Lab to the After Hours feed, made the same case from a 20-year software-engineer’s perspective. His framing:

“Being able to look at a piece of software you’re using right now — let’s say Premiere — see something that doesn’t exist, and in real time build a plugin and drop it in… that’s wild to this code monkey.”

He paired the demo with a sober coda: the engineering mindset still matters. Vibing up a prototype is one thing; understanding how the architecture under it actually fits together is what keeps the thing alive past day three. His pragmatic advice: dig into what the AI built for you, and decide whether it made good decisions.

He also mentioned rebuilding what a former personal assistant used to do, in a day, using AI plus an automation layer (an n8n-style glue framework). Faster, more consistent, no inbox triage.

ClaudePremiere pluginsn8n-style glueReal-time iteration

The throughline

None of these were demos of a finished product. They were people describing what they had built this month, in evenings, between shows. A control-room utility, a community affiliate engine, a Premiere plugin that didn’t exist that morning. The common stack was small: a code-aware AI in the loop, a hosted runtime (Vercel or local), a Git repo, and the discipline to audit keys before going public.

Part Two

The Other Technical Threads That Day

While the main stage cycled through gear advice and grant strategy, the back channel on After Hours wandered into seven or eight different rabbit holes. The notable ones:

Vercel security incident — rotate keys and any linked GitHub tokens

Mid-morning, a remote attendee in Wisconsin flagged a freshly announced Vercel security incident, with the public-service reminder: if you have keys hosted on Vercel, rotate them now — and rotate any GitHub tokens connected to it.

The room responded with a useful self-audit pattern, particularly relevant if you’re about to flip a private repo to public:

  • Run a Claude-driven security pass over the codebase before flipping visibility.
  • Hook MCP into the connected services so the audit can confirm what’s actually wired where, not just what’s in .env.
  • Where possible, replace static keys with provider-to-provider auth (e.g. Supabase ↔ Vercel integration) so there’s no long-lived secret to leak.
  • Where you must use a key, set it to auto-rotate.
06:38 PM CDT · Creator Lab Q&A

Comfy UI inside Unreal Engine — with Cesium streaming the world

One of the more striking demos: a plugin that brings the Comfy UI node graph into Unreal Engine, on top of Cesium tiles streaming Google-Earth-style 3D context. You can frame a camera anywhere on the planet, then let a local generative model retexture the scene photorealistically — without anything leaving the laptop.

The plugin was built by a recently-ex-Epic engineer named Michael; it was posted to GitHub a few days earlier and announced on LinkedIn. The companion plugin from an earlier demo (using Nano Banana to retool camera context) was called Ugen.

Unreal EngineComfy UICesiumOpen source / GitHubLocal generative
06:12 – 06:21 PM CDT · Creator Lab demo

Local AI is closer than you think

The same talk demoed a fully local image-generation pipeline running on a single laptop — a deliberate choice given the conference Wi-Fi. Models referenced:

  • Z Image — local model generating photo-real full-HD frames in seconds.
  • LTX 2.3 — local video generation, ~5-second clip in a few minutes on the same laptop.
  • Google Genie 3 — real-time interactive video generation; cited as “a real-time video generative platform that can act like a game engine.”
  • Apple Sharp and NVIDIA Lyra — both for Gaussian-splat generation; Sharp is fast (a few seconds), Lyra is slower but generates large worlds.
  • Chaos Vantage with NVIDIA DLSS — real-time, full-frame ray-tracing onto an LED wall, sourced from traditional 3D apps (3ds Max, Maya, Houdini), bypassing the Unreal-only path.

The forward-looking take from the room: the cloud will always have more compute, but local has crossed the threshold of useful for production, and the trajectory only points one way.

06:59 PM CDT · Late After Hours, side conversation

Flask · Gunicorn · nginx — the small-Python stack, explained

One of those short evening detours that earns its keep. Someone asked what flask, gunicorn, and nginx actually were and how they relate. The answer that landed:

“Flask is to Python what a hand-rolled cigarette is to a vape. Fewer parts. You choose what goes in. It works the same way it always has.”

The three-piece breakdown:

  • Flask — the minimal Python web framework you write your app in.
  • Gunicorn — the production-grade web server that accepts HTTP connections and feeds them to Flask.
  • nginx — sits in front of Gunicorn, handles SSL, serves static files, proxies dynamic requests through.

Three pieces, each doing one thing — and a useful mental model for anyone who’s been staring at gunicorn -b 127.0.0.1:8903 in a tutorial without quite knowing what each layer was for.

10:15 AM CDT · Office Hours follow-up

vMix render-time rule — stay under 16 ms a frame at 1440p60

A Saturday audio-sync glitch during the show was traced back to vMix being overtaxed at 1440p60. The room-shared rule of thumb: watch render-time per frame; you need to remain < 16 ms. Saturday opened in the 12–14 ms range and drifted into the 20 ms range — every frame past 16 ms eats into the buffer until sync slips and frames drop.

10:19 AM CDT · Office Hours buying-cycle question

Mac Studio cycle — M3 Ultra now, M5 Ultra perhaps in October

An eight-week order delay on an M3 Ultra Mac Studio led to a buying-cycle dive. The Mac Rumors buying-guide framing was cited (last release March 2025, average cycle 547 days, currently at day 410+). One panelist summarized the Bloomberg/Mark Gurman read of a possible October refresh, attributed to component-supply pressure across the industry.

The room’s consensus on now or wait: if it’s a money-making machine, buy it; the 14-day return window at the Apple Store gives you a real-world load-test escape hatch.

08:15 PM CDT · Late After Hours

Starlink — reliable enough for production, but it reboots when it wants to

A panelist who recently moved to a rural property reported Starlink had been “rarely a noticeable hiccup” for daily Zoom hosting. The one operational gotcha worth knowing: the dish takes firmware updates on its own schedule (somewhere between 02:00 and 04:00 in his time zone) and reboots without asking. Unlike Windows or macOS, you can’t defer it. Pick a time-window setting that matches when you’re definitely not on a stream.

What the day rhymed on

The shape of the day, looked at from the back of the room: every conversation kept arriving at the same shoreline. Smaller teams, faster feedback loops, more capability per laptop, fewer reasons to wait for someone else to build the thing you need. The After Hours version of that was less triumphant and more matter-of-fact: here’s the gear app I built last night, here’s the security note from this morning, here’s the three-line stack you’ve been quietly running for years.

That’s the room. That’s the party.