Newsroom, our client: Sizing up AI’s role in illuminating journalists’ daily routine

Problem is

Spending an exorbitant amount of time rifling through countless press releases and crafting concise yet compelling summaries that adhere to a single style.

Possible clue can be

The automation of gathering, analyzing, and summarizing news stories by involving AI technologies.

Oxagile’s first move

Our team took the initial step by crafting a PoC based on the OpenAI API. Its successful implementation paved the way for the development of a fully functional AI solution.

How did the news aggregation solution further evolve?

  • Reaching 100% product uniqueness

  • Drawing LLMs in tackling the challenge

Oxagile managing LLM use: How we dealt with the task

Falcon LLM got the nod

Through the utilization of the Project Gutenberg dataset and the works of Shakespeare, we have successfully fine-tuned the LLM to the point where it can precisely mimic any defined text style.

Hugging Face

Oxagile’s team opted for the usage of the Hugging Face infrastructure to train the large language model.

PEFT method

A Parameter-Efficient Fine-Tuning (PEFT) approach allowed us to fine-tune small subsets of the model’s parameters, freezing the original pre-trained weights of the LLM.

Performance check

We utilized the mathematical method of evaluation, first making the words comparable through coding, and then identifying the threshold level of being “Shakespearean enough”.

Why AL algorithms surpass other methods in this case: Three key benefits of LLMs

Here’s another AI story for you: Almost 100% sound detection accuracy achieved

First things first, Oxagile figured out which ML algorithms would work best for classifying lung sounds and become the cornerstone of a smart stethoscope solution.

Once the algorithms were sorted, it was up to our team to make sure we could accurately identify different sound classes. We were determined to push the limits of the neural network’s accuracy, so here’s what we did:

  • Came up with a roadmap that included a few Explainable AI tasks to perform;
  • Dug deep into all the possible reasons why two sound classes were being misinterpreted.