Impel’s mission to bring safer, more effective and more contextually aware artificial intelligence for the auto industry has taken two giant steps forward, the company said.

The provider of automotive artificial intelligence technology announced the public showcasing of its domain-tuned, production-grade large language model and the release of Archias, a research-first expert model designed to advance AI safety standards across high-context verticals.

Impel’s custom-built verticalized LLM, optimized in partnership with Amazon Web Services for the automotive retail environment, was recently featured as a case study at Meta’s LlamaCon Conference.

The model, selected by AWS for its performance and business impact, was recognized for delivering faster, safer and more contextually accurate AI and for tangible business value in real-world production settings in automotive retail workflows.

Impel replaced its existing third-party LLM with a fine-tuned Meta Llama model deployed on Amazon SageMaker, achieving a 20% improvement in accuracy for customer-facing automotive applications. Impel said the deployment highlights its commitment to building LLMs that solve real business problems.

Impel said it paired research with production as its research and development team authored a research paper introducing Archias, a domain-specific expert model designed to detect and protect against prompt injections, malicious queries and other adversarial attacks.

Released alongside a benchmark dataset and evaluation framework, Archias serves as a research initiative to accelerate understanding and improvement of AI safety in automotive and other specialized domains, the company said.

In benchmark testing, integrating Archias with open-source and proprietary LLMs improved model output accuracy by 3.6% to 20.7%, depending the model. Impel said those results demonstrate “significant advances in security and domain relevance across multiple architectures.”

The paper has been published on arXiv, an open-access research-sharing platform, and is under review by scientific publications. It proposes a new model for identifying and mitigating threats in AI systems used in high-stakes industries.

“General-purpose AI can’t meet the demands of a high-stakes, high-context industry like automotive,” Impel co-founder and CEO Devin Daly said. “This research is not just a technical milestone, it’s a signal that purpose-built vertical AI is essential to doing business in complex industries.

“With this release, we’re giving the entire industry a blueprint to help ensure the long-term health and advancement of AI in automotive retailing.”

As a research project, Archias is designed to foster a transparent, collaborative approach to AI safety and alignment through evaluation and learning. The framework evaluates key threat categories, including in-domain misuse, adversarial prompts and pricing manipulation, and reflects real-world risks faced by dealerships and OEMs.

Impel said the strongest safeguards in all industries require models that not only reject out-of-domain and adversarial requests, but also understand the workflows and risks unique to the industry they serve.

“By releasing our methodology and benchmark to the public, we’re encouraging the industry to take a research-driven, transparent approach to AI safety, rather than relying on quick fixes or hoping for silver bullets,” said Dachi Choladze, Impel’s chief innovation officer and co-author of the research paper. “As generative AI becomes embedded in customer-facing workflows and industry-specific applications, the cost of errors or misuse rises exponentially.

“General-purpose models aren’t enough. We need AI systems that are adaptive, responsible and deeply aligned with the domain they serve. By sharing our research, we’re helping the industry move forward together.”

The company said both the LLM and Archias build on the Impel AI Operating System, an enterprise-grade automotive AI platform purpose-built to serve every dealership department, from sales and service to marketing and merchandising.

The platform blends foundational models from providers like OpenAI, Anthropic and Meta with Impel’s proprietary verticalized models and is designed to deliver “real-time performance and domain-specific alignment at scale,” the company said, noting it has been involved with more than 33 billion consumer interactions and $8 billion in influenced revenue for more than 8,300 dealers in 53 countries.

“These parallel achievements — our production-grade domain-tuned LLM and the Archias research initiative — show why industry-specific verticalized AI, backed by rigorous research, is the path forward for safer, more effective enterprise deployments,” Daly said.