Business case

• Automate invoice validation by using LLM for OCR extraction
• Once the data is extracted, call external functions to enhance the data and validate the company
• Document Intelligence and Custom Vision are not having a good result when checking random invoices

Points of improvement

• Add parallellism using asyncio/httpx instead of requests
• Add AAD authentication if available
• Add metrics comparison

Main differences

Anthropic

Perks

• Easiest to use
• Don't need to inform a region (the API Gateway takes care of the request)

Disadvantages

• Only 2 SDKs (Python and Typescript)
• Lacking topic of discussions on the Internet

Google Gemini

Perks

• Easy to use
• They manufacture their own chips (easier to scale up)
• Creators of Transformers and self-attention approach

Disadvantages

• Lacking topic of discussions on the Internet
• System instructions not supported for vision (???)

Microsoft Open AI

Perks

• More material on the internet (StackOverflow, Medium and other sources)
• Have integrated options like "Add you own data"
• CSS of Microsoft is the best on the planet
• Extra layer of security among APIs for the rest of the environment
• Pioneer in the LLM area (together with Open AI)

Disadvantages

• More complex to use

PoC Results

Data quality

Invoice	GPT4Vision	GPT4o	Gemini	Claude
1	2nd	2nd	4th	1st
2	2nd	2nd	4th	1st
3	2nd	2nd	1st	2nd
4	3rd	1st		1st
5	1st	4th	1st	4th

Time to process all invoices

GPT4Vision	2:12
GPT4o	1:49
Gemini	1:04
Claude	1:11

Podium

	Model
1st	claude
2nd	gpt4o
3rd	gpt4vision
4th	gemini