During the Agility Recovery training course project, Asymmetric built a custom AI knowledge system trained exclusively on Agility Recovery's internal documentation. Rather than relying on general-purpose AI outputs, the team ingested 197–277 source documents to create a domain-specific expert system capable of answering questions, generating content, and producing learning assets in Agility Recovery's own language and terminology.
This approach directly improved the quality of the [1] RISE 360 training course by grounding all generated content in the client's actual materials rather than generic industry knowledge.
The first GPT instance hit a memory ceiling at approximately 100 documents. The team addressed this by:
1. Creating a second instance on a platform with higher capacity
2. Accepting the current corpus as sufficient for the initial course build, with a plan to expand as new documents become available
"We felt like it was pretty bloody smart. So we didn't really need to get a whole lot deeper." — Mark Hope
The custom GPT system supports a range of content generation and analysis tasks:
| Capability | Description |
|---|---|
| Q&A / Research | Answer questions about Agility Recovery products, terminology, and services |
| Mind Mapping | Generate visual concept maps from the full document corpus |
| Quiz Creation | Produce knowledge-check questions grounded in actual content |
| Report Generation | Summarize topics or products across multiple source documents |
| Flashcard Creation | Build study aids from document content |
| Block Analysis | Analyze RISE 360 content structure (e.g., text block ratios, interactive element distribution) |
| Course Content Drafting | Generate module text using client-specific language and product names |
The initial course content draft was criticized by the client for being too cybersecurity-focused and not accurately reflecting Agility Recovery's core product language. The custom GPT addressed this by:
"It's not just taking ChatGPT garbage... it's taking 197 Agility Recovery documents and trying to understand them all and then help create the course from that." — Mark Hope
One concrete output of the GPT-assisted analysis was a RISE 360 block distribution recommendation:
Overuse of interactive elements (flip cards, drag-and-drop) was identified as a risk — making courses feel "childlike" rather than professional. The GPT helped audit existing modules against this ratio.
Example block summary from Modules 1–3:
- 42 text blocks
- 22 knowledge checks
- 15 interactive tabs
- 12 accordions
- 12 statement blocks
- 8 quotes, 1 image, 1 gallery
The system is designed to grow with the project: