 on a different project in this case related to medical prescriptions. Our client is a multinational company that operates in South Central and North America and monthly receives thousands of images and its operators manually transcribe them coding each entity according its database codes. For each prescription is necessary to code the doctor, the name of the doctor, the medicine with its information, dosage, application and laboratory and the institution where it was prescribed. You must know that detecting names is not enough, entities must be code in an exact way. Additionally, it's important you know that images are generally bad, low quality, tangled out of focus, shaded, mixed with other images taken from too far, etc. Here we can see an example, we can see the name of the doctor in blue circle and in red, medicines, dosage and form of application. Here we can see two prescriptions, another one and the last two. Notice that the prescription is inside the picture we have, okay? So it's very difficult to detect exactly where the prescription and the entity is on. Okay, which is the objective? The objective of this company obviously is to reduce transcription times and the cost. So we have set out to preprocess each prescription in order to present the obtained results on screen with the proposal of validation and eventually correction by an operator. The project is economically viable only if operator intervention is minimal allowing them to process more prescriptions by the hour. We have three problems, three issues to face. The first one is the transcription. The second one is the entity recognition. This implies detection of the entities in the text. And the last problem we have is coding the entities, that means assign each entity its code in the basis of master tables. Okay, for the first problem, we tried, first off, we tried with public models in particular transformant Titan model, but we decided to create our own HDR and baseline model with almost 10,000 prescriptions. We obtained better results. The second problem with respect to name entity recognition, we analyze three different options. The first one was to use NLP libraries like Spacey, NLTK, etc. The second one was to use cloud services like Amazon Comprehensive Medical and others. And the last one was to use transformer models, specifically large language models. You know, these models require not training because understand context and generate coherent responses. So this last option has been the option of choice. So we tried two models, a paid for and a free one. The first one, GPT 3.5 Turbo and Lama 2. Using GPT, we had superior, it was superior to Lama 2, at least in our test. And the follow with procedure. It's important here to have a sweet table prompt engineering and the implementation of retrieval augmenting generation. Starting from a grand through with the format of 1000 target descriptions saved in a vector store upon processing each prescription look up of the closest format in vector store and use it as one shot example. And the use of embeddings allow us to increment the percentage of recognized recognized entities. Here we can see a typical JSON file we are obtaining after the process with with each entities. And the last, the last phase is the coding. It is necessary to normalize the JSON file. So we are doing a general cleaning normalization of approbations orthographic correction Levenstein application, etc. Here we can see a before after example of the process, you can see the original text and below the normalized. So, addition, once normal is normalization is accomplished. Accomplished matching with database is relatively simple. Additionally, we're testing right now embeddings to search by cosine similarity when entities doesn't match exactly results until now. 90% 95% is detected and correctly code. And with respect medicine 70% is detected and correctly code and 20% is partially detected. Obviously, we are working right now to increment the percentage of efficacy improving the improving the process in all its stages. Right now, right now we are in a full concept testing with the client. The test consists in the preprocessing of 50,000 prescriptions and the generation of metrics that can measure productivity. That's all thank you very much for your attention. Are there any questions. Do you get a confidence score with detections and can you use that. The customer is very happy with the results we are getting right now. The percentage of the pre-op 90% 95% in doctors and in medicine 70%. But we're trying to improve the process. I know we can improve the process in each stages. We have a lot of work to do we can in increment the percentage. Welcome.