Autonoma, a leading loT platform for industrial machinery manufacturers, faced the critical challenge of addressing equipment downtime, which can be costly and negatively impacts their clients. To enhance technical support, Autonoma aimed to develop an Al agent that provides troubleshooting instructions for complex technical issues. However, the process of building and verifying this agent was anticipated to take several months.

Before partnering with Vijil, Autonoma encountered several challenges:

• Technical limitations: Early prototypes had issues with accuracy, including generating hallucinated responses and failing to switch between languages(German and English) as peruser request.
• Security risks: The initial prototype was susceptible to misuse and attacks.
• Protracted development timeline: Developing custom testing, evaluation and guardrails would entail extended development cycles.

In addition, the team had decided that the agent would run on DigitalOcean's GenAl platform because of ease of use and low-touch integration easily into their existing infrastructure.

Autonoma partnered with Vijil to accelerate the agent's development and enhance its reliability. The collaboration focused on several key improvements:

1. Model optimization

The team identified that their initial model of choice generated a high rate of hallucinations when deployed with a Retrieval Augmented Generation (RAG )function connected to the firm's knowledge base - and was vulnerable to prompt injection attacks.

Vijil security and safety testing pointed to a model that performed less well in the inclusion of images, but that would significantly lower the risk from simple agent misuse via prompt injection as well as jailbreaks.

2. Accuracy enhancement

Vijil reorganized Autonoma's online help content, ensuring it was accessible in both English and German, thus improving retrieval accuracy for RAG functions connected to the new model - and reducing the rate ofhallucinations.

The English and German content was separated into different knowledge bases to improve the ability of the agent to cite content in the user's language.

3. System prompt refinement

The system prompt was expanded to include strict instructions, improving agent responses and preventing misuse.

Explicit guidelines allowed the agent to alternate accurately between languages.

4. Guardrails implementation

Vijil integrated safety policies and misuse prevention instructions, enhancing the trustworthiness of the agent, preventing the agent from being exploited or performing tasks outside its intended scope.

Users can now enforce these policies through guardrails deployed on the Digital Ocean platform.

Working with Vijil, an expert Al agent builder on the DigitalOcean Gradient Platform, the Autonoma team deployed a secure, reliable agent in just one week, saving a lean Autonoma team months of development time and costs. Using the Al agent, Autonoma has reduced their customer's time to troubleshoot each incident and is poised to scale their operations faster to more customers.The collaboration led to significant changes in the development and deployment process, achieving remarkable outcomes:

Rapid deployment

The AI agent was quickly developed and deployed in just one week, a dramatic reduction from the previous timeline.

Improved accuracy

The agent now generates reliable content with adequate references and grounding in original documentation, along with successful handling between languages.

Enhanced security

New measures made the agent resilient to common prompt injection attacks, increasing safety for end-users and reducing the risks of deployment in production.