Document Forge

Over a year ago, customers started asking for something we didn't think would be difficult. We certainly didn't think it would take more than a year. The longer we worked the problem, the more impossible it felt.

AuditionAI should generate PowerPoint decks as good as one made by an experienced professional.

We made real progress early. Our biggest win was allowing users to upload a template, which gave us consistent, repeatable output. That solved the "we need a deck that matches our brand" problem.

Our early templates approach produced genuinely good work — well-structured, branded correctly, internally coherent. But good and human-level creative are not the same thing.

Why weren't we satisfied?
It didn't solve creativity.
It didn't solve iterative design improvement.
And it definitely didn't solve taste.

We poured over a thousand hours into the problem.

The Deceptive Simplicity of PPTX

At first glance, PowerPoint seems simple. A PPTX file is just XML. Any LLM can generate XML.

But that's not where the real challenge is.

Layout is fixed. There is no vertical overflow to catch your mistakes. A word too many and the whole slide breaks. A font size up and columns collide. Every pixel carries weight.

That realization led us to what felt like a genius plan a year ago: have the AI create the deck, render the deck into screenshots, feed those screenshots into a review loop so the AI can see what actually happened and improve it.

Simple, right?

Not even a little.

The Constraint That Changed Everything

Lab tests looked promising. But because our platform runs in our customers' cloud environments, we don't fully control the runtime. That made the rendering and review loop much harder than it looked on paper.

We studied how every major AI platform handled PPTX generation. A year ago, most of the results from almost every App we tried underwhelmed. This failure to impress was so widespread we became convinced that if we could solve this well, it would be a real differentiator.

When Claude Shipped Our Idea First

We kept an eye on the rest of the market. Then one day we noticed Claude had made a major leap in PowerPoint quality.

A quick review made it clear: they were doing almost exactly what we had envisioned.

That observation also confirmed we were on the right path. The difference was that we couldn't simply copy the approach — we had to make it work inside arbitrary customer cloud (or on prem) environments.

What Finally Cracked It: The Four-Layer Toolchain

What ultimately cracked it was noticing the newer toolchain behind Claude's improvement. Understanding the pattern and the opensource libraries they choose inspired us to double down efforts and reach a worthy conclusion.

Proof of Parity

AuditionAI

Claude

The Insight That Changed How We Think About Latency

There was one unexpected discovery that reframed our thinking on the entire feature:

A thirty-second wait feels frustrating when the output is mediocre. The same thirty seconds feels like magic when the result lands exactly right.

Document Forge: The Bigger Picture

At the same time, our team had been working on a broader architectural shift inside AuditionAI that expands on our platform's ability to safely write and execute code. We wanted something much more powerful. Stay tuned in a week or two for an update on this exciting development!

We called this a primary quest internally — because it doesn't just solve PowerPoint. It unlocks a whole category of future capabilities we've been waiting to build.

The Saturday Call

We were only weeks away from finishing when I got a Saturday call from an executive at a new customer. He was kind, but direct:

Internal Reveal

Just days ago, the team ran a live internal showcase using the same prompt in Claude and AuditionAI.

In my humble — and hopefully objective — opinion: the results are now aligned.

AuditionAI Outcome

Document Forge generating a robust, detailed PowerPoint presentation end-to-end inside AuditionAI's platform.

What the QA Loop Actually Does

The magic of Document Forge isn't just that it generates decks. It's that it sees what it generated and improves it. Here are real examples from the development process: first drafts on the left, AI-corrected versions on the right.

Example 1Text Overflow Fix

First Draft

AI Vision: QA Corrected

First draft: Text runs outside the safe zone → Corrected: Text resized and repositioned

Example 2Layout Alignment

First Draft

AI Vision: QA Corrected

First draft: Uneven spacing throws off visual balance → Corrected: Proper alignment and spacing applied

Example 3Text Overflow in Content Area

First Draft

AI Vision: QA Corrected

First draft: Bullet points overflow in multiple areas → Corrected: Font sizes and spacing optimized across all content zones

This feedback loop is the difference between "technically valid PowerPoint" and "this actually looks professional."

One thing I think is especially cool: the model we used for this in AuditionAI wasn't even an Anthropic model.

That said — Sonnet and Opus are still some of my favourite models when I want strong creative and design instincts. They're excellent when the goal is to build something beautiful.

What AuditionAI Actually Is

For those who don't know us well: AuditionAI is the interface and orchestration platform enterprises use to deliver GRC-first AI.

We're not a replacement for Anthropic or OpenAI models. We are the control layer between users and those models — the layer that helps enterprises govern, observe, and protect company data and intellectual property, regardless of which model is doing the work.