Improving AI accuracy and refining analysis approach

What I worked on today

• Spent the entire day figuring out how to improve AI accuracy.
• OpenAI’s screenshot analysis was unreliable, requiring an alternative approach.
• Tested Amazon Textract, Amazon Rekognition, and Google Vision—ultimately chose Google Vision for better parsed text extraction.
• Implemented Google Vision before AI analysis and refined the OpenAI prompt.
• Realized that the prompt is now heavily focused on SaaS pricing, making it effective for that niche but requiring custom prompts per use case.

Lessons learned

• AI models need structured input—Google Vision provides better parsed text, improving analysis results.
• Different AI tools work better for different use cases.
• Frustration is part of the process, and I nearly gave up but glad I pushed through.

Challenges faced

• Spent all day testing different AI solutions.
• Needed to refine the prompt heavily to focus on SaaS pricing.
• Future AI use cases will require different prompt structures.

What’s next

• Consider fundamentally changing the way the platform processes AI analyses.
• Decide whether to launch first or restructure AI use cases before launching.
• A potential customer has a different use case than the current AI analysis, so evaluating adaptability.