Neuromnia is leveraging Meta's Llama 3.1 language model, to transform Applied Behavior Analysis (ABA) therapy for individuals with autism spectrum disorder.The company has developed Nia, a human-centric AI co-pilot that aims to enhance clinician productivity and improve access to care for individuals on the autism spectrum.
Jay Gupta, Co-Founder and CEO of Neuromnia, highlighted the impact of Llama 3.1 on their mission. "After testing with Llama, we were impressed by its state-of-the-art performance in natural language processing tasks," Gupta said. "Combined with our CPO's clinical expertise, Llama has enabled us to build a comprehensive dataset that will help reduce the administrative burden for clinicians and provide quality care."
Neuromnia's journey with Llama 3.1 began during the research and testing phase of product development. The team started working with the 70B model, quickly recognising its potential to support their goals. They fine-tuned the model using a curated dataset compiled by Josh Farrow, Co-Founder and Chief Product Officer, who is a Board Certified Behavior Analyst (BCBA).
The platform is designed to address ABA workforce shortages by automating key elements of treatment planning, documentation, and modification. Nia suggests major treatment components such as goals, intervention strategies, and procedures, allowing clinicians to manage larger caseloads more efficiently. With autism affecting one in 36 children, Neuromnia's solution addresses pressing challenges in autism care.
Gupta emphasised the importance of open source in their development process. "Open source has empowered us to build and scale our solutions without succumbing to vendor lock-in," he noted. "The community support has been a welcome source of information and advice for LLM creation and deployment."
While integrating Llama 3.1 into their platform, Neuromnia faced some early misconfigurations and bugs. However, the team resolved these issues through trial and error. To enhance the model's output, they incorporated techniques like prompt engineering and retrieval-augmented generation (RAG).
Looking ahead, Neuromnia plans to continue using Llama, evaluating each new release to ensure their platform remains cutting-edge. "While out-of-the-box LLMs can sometimes struggle with technical or complex tasks, we've significantly reduced error rates through a combination of prompt engineering, RAG, fine-tuning, and techniques like semantic search," Gupta explained. "These advancements help ensure that our solution will continue to deliver accurate, hyper-personalised results that make autism support accessible to every family."