I built a working AI video engine where agents write and direct, code handles timing and quality, and a person approves instead of editing. That's the system Elva is hiring for, and I've already shipped one.
Every item in this section is in the working code today. I checked each one against the source before writing this page.
Every uploaded picture is read once by a vision model for its text and a one-line description. The director places the invoice screen on the line about invoices, not in upload order.
Each on-screen card is anchored to the moment its word is actually spoken, measured from the voice. Re-record the voice and the cards re-time themselves.
Templates, thresholds, voices, palettes and sound effects are versioned records. A new style is a new version, not a new editor.
Approvals are tied to the exact inputs. Edit one line upstream and only the pieces that depend on it lose their approval.
Paste a URL: the page becomes a dated snapshot of facts, and the brand palette is read from the site's own stylesheets for a person to confirm.
A missing or failed picture renders as a clean fallback card. A model that ignores a rule is asked again, then the rule is applied for it.
One film through every gate in about six minutes, including a rule failing on purpose and an approval going stale.
Elva wants a system that takes an idea and references, analyses them, builds a narrative, picks visuals and audio, and outputs a structured video. Here's where I've already built each piece.
At 100 videos a day, nobody can check every card and scene by hand. So every version has to pass the machine checks before it reaches me, and I give the final verdict as the approver, not the editor.
They run on every version of every storyboard. Any error blocks that scene's approval and the final render, and the final render checks again and refuses to start if errors remain. Six of the rules are warnings that flag work for a closer look.
Free disk space, type checks on the app and the engine, the full test suite (661 tests), a scan for hard-coded off-brand colours, a repeatability check, a clean build, and a 6,000-frame endurance render.
Approvals are tied to the exact inputs. Change one line and only the pieces that depend on it need approving again; everything else keeps its sign-off.
Two working projects, combined into one system built for volume. And it adapts: every client has their own workflow, formats and approval steps, and the engine is built to be customised to the pipeline you already run.
The version 3 plan is a set of building blocks, not a fixed product. Each one can be shaped around your needs:
I'll review it and give you the exact build time, live in a demo or right after. That includes building from scratch if it doesn't fit my current engine.
I design and build AI production systems end to end: the agents, the rules that keep them honest, the rendering, and the review screens. I'm based in the Philippines and work European hours.