Top Gen AI Demos of AI Applications With MLRun

Gen AI applications can bring invaluable business value across multiple use cases and verticals. But sometimes it can be beneficial to experience different types of applications that can be created and operationalized with LLMs. Better understanding the potential value can help:

• Garner excitement and collaboration across the organization
• Help secure resources
• Support planning strategies, from a single use cases to many and from PoC to operationalization
• Drive more ideas for innovation
• Help de-risk pitfalls
• And more

In this blog post, we’ve curated the top gen AI demos of AI applications that can be developed with open-source MLRun. Each of these demos can be adapted to a number of industries and customized to specific needs. Follow along and choose the most relevant ones for your needs.

You can also watch the complete library of demos here.

1. Smart Call Center Analysis Application

Build a call analysis platform for call centers. The application analyzes customers calls and generates actionable insights for agents, management and downstream applications. It helps support agents, create tailored recommendations for customers, and more. This improves the customer experience, helps optimize first call resolution, enhances operational efficiency, supports decision-making and helps meet compliance regulations.

The gen AI application is built on a multi-step pipeline that includes diarization, transcription, PII filtering, analysis, post-processing, among others. Output structured data is stored in a database, accessible for reporting or downstream applications. Visual dashboards provide metrics like topic summaries, ag and data filtering outcomes.

Open-source MLRun automates the entire workflow, auto-scales resources as needed, and automatically logs and parses values between the different workflow steps. Watch the smart call center analysis app demo.

2. Monitoring and Fine-Tuning a Gen AI Banking Chatbot

Learn how to monitor and maintain the integrity of a banking chatbot, ensuring it adheres to banking domain rules. This means restricting responses to only banking-related topics. For example, a query like "What is an artificial neural network?" is flagged as unrelated.

Fine-tuning the chatbot ensures it meets industry compliance regulations, prevents user frustration by going off-topic or incorrect responses and improves overall satisfaction and engagement. It also prevents “abuse” of the model by users, who might use it for purposes not intended by design. The feature is particularly relevant in industries like banking, healthcare and legal services, where precision and adherence to guidelines are non-negotiable.

Model responses are evaluated with an LLM-as-a-judge. Poor accuracy triggers an automated feedback loop. This automated workflow fine-tunes these poorly-performing models with new training data and redeploys the fine-tuned model. This ensures continuous performance improvement while supporting scale across multiple models and use cases. 

Watch the fine-tuning demo here.

3. Real-time Gen AI Co-Pilot (Wealth Management Use Case) 

See how a gen AI co-pilot amplifies a client relationship management scenario in private banking, emphasizing personalized service, strategic investment advice and proactive support for the client’s needs. In the presented scenario, the banker identifies and recommends a relevant investment opportunity based on the client’s history, while proactively sharing research materials from reputable sources to support informed decision-making. This fosters a sense of trust and expertise while generating more business for the bank.

In addition, the co-pilot helps the agent to anticipate future opportunities, like biotech investments, based on client interests, which expands the bank’s role in the client’s portfolio. They also offer complementary banking services, including opening a student bank account with tailored benefits for the client’s daughter's relocation to the UK, which reduce client stress and enhance the relationship.

The co-pilot helps the human agent provide personal touches, such as acknowledging the client’s daughter’s achievements and offering tailored solutions, to build trust and loyalty. This long-term retention through proactive service ensures steady revenue from high-net-worth clients. Finally, the co-pilot created hyper-personalized email generation based on the conversation (to be sent as a follow-up after the call) for accountability and to close the deal.

Watch the wealth management co-pilot demo here.

4. Real-time Insurance Gen AI Co-pilot (Luggage Claim Use Case)

In this second co-pilot example, the gen AI agent helps the human agent guide a customer through reporting and resolving a lost luggage claim. The agent provides customer details, analyzes call sentiment, summarizes the call and suggests actions for the human agent to take, based on company policies. Suggestions are updated as the conversation evolves and the customer provides more details.

For example, the gen AI agent recommends checking policy coverage and requirements. explains that cash and jewelry are not covered in checked luggage and retrieves the insured amount, per the client’s request. These actions help reassure the customer and reduce uncertainty and frustration.

Watch the insurance co-pilot demo here.

5. Hyper-personalized Banking Gen AI Agent with MongoDB

See how a gen AI banking agent provides clients with hyper-personalized recommendations for credit card choices. The agent analyzes personal data like the client’s income, credit rating and spending habits, improving relevance and satisfaction. Hyper-personalized suggestions are generated to encourage clients to choose products aligned with their needs, driving conversions. This also makes upselling opportunities more effective. In addition, the conversation tone and style are adapted to the client’s profile (e.g., casual for younger clients, professional for older clients), making the interaction feel engaging.

The solution is based on MLRun and MongoDB. The workflow, orchestrated by MLRun, retrieves structured client and credit card data (e.g., income requirements, fees) from the MongoDB cluster in real-time. This is used to hyper-personalize the offering and conversation. It enhances raw data with AI-generated descriptions and conversational responses. It also leverages machine learning to learn from previous interactions, further refining personalized recommendations over time.

Watch the hyper-personalized agent demo here.

6. Multi-Agent Banking Chatbot with NVIDIA NIM and MLRun

See a multi-agent banking chatbot that enhances customer interactions by operating multiple agents: a loan agent for loan inquiries, an investment agent for investment opportunities and a general agent for customer service.

The system accurately classifies user queries to direct them to the correct agent, improving efficiency, reducing operational costs and enhancing customer satisfaction with instant, accurate responses. For example, a query about buying a house financing correctly routes to the investment agent.

The chatbot is built using open-source MLRun, LangChain and NVIDIA NIMs. NIM is deployed with MLRun as serverless functions with minimal setup and API keys are securely managed as secrets for production readiness. MLRun’s built-in LLM gateway is a solution allowing for model modularity and monitoring for multiple models and use cases. LLM-as-a-Judge monitors the classifications made in order to choose the correct agent. Sessions and conversations are logged for future fine-tuning and compliance.

The solution is designed to be scalable and modular. It enables seamless handling of multiple concurrent users and flexibility in model switching and reusing classifiers, prompt templates, and workflows, to optimize performance and costs.

Watch the multi-agent banking chatbot demo here.

7. Customer-facing Gen AI Agent for a Jewelry Retailer

See how an AI-powered shopping assistant can assist new customers with complex purchasing decisions, even in scenarios where minimal user information is available. AI can mimic the behavior of a skilled salesperson, providing personalized recommendations and seamless navigation through the shopping experience.

In the scenario in the demo, a husband shops for jewelry for his wife. Despite the lack of prior purchase history, the assistant efficiently guides him through the decision-making process despite limited inputs, like the customer specifying only the occasion and preferences (e.g., dislikes rings, bracelets, and uncertainty about earrings).

If there was a purchase history, the agent would have recommended items similar to previously purchased ones, and would avoid suggesting items that were already purchased.

The agent narrows down options based on the customer's evolving preferences. Then, the assistant suggests jewelry pieces (e.g., gold necklaces) without overwhelming the customer with endless questions. Suggestions are tailored, such as pink gold necklaces when prompted for a specific style. The agent also answers practical questions about payment methods (e.g., compatibility with American Express) by retrieving relevant policy details and directs the customer to the product page for more details. This makes the shopping journey intuitive and engaging while the customer feels supported. From a business perspective, the path to purchase is accelerated and customer loyalty grows.

The agent can scale to handle multiple customers simultaneously in real-time, offering personalized support without requiring human intervention. It also learns from conversations to improve future interactions for repeat customers.

Watch the gen AI agent for retail here.

8. Customer-facing Retail Gen AI Chatbot

In this second chatbot example, the chatbot assists customers in a fashion e-commerce setting by creating personalized fashion recommendations, like suggesting items based on customer input (e.g., color, occasion, or outfit preferences) and adapting responses dynamically to customer preferences during the interaction. The chatbot is also able to answer questions about store policies like shipping and returns. All responses align with the company’s tone and code of conduct. The chatbot is designed to communicate in a friendly and professional manner, increasing customer satisfaction and conversions.

The chatbot is built on a reusable, modular architecture based on open-source MLRun. Components include session loading, query refinement, guardrails (toxicity filter and subject classification) and history saving, and they are reusable for multiple applications.

After being prompted, historical sessions are loaded for query refinement. After passing through guardrails, everything passes through the application logic, which is also where LangChain and LangGraph reside. It is deployed as a scalable API, supporting seamless integration into production environments.  

The demo uses the Mistral-7B-Instruct-v0.2 model for its balance of quality, cost, and performance. Input and output data are collected for RLHF fine-tuning of the LLM. Fine-tuning focused on reducing prompt size, embedding guardrails, and improving response quality. Crowd-sourced evaluation compared original vs. fine-tuned models, leading to a 71% preference for the new model.

Watch the retail gen AI demo here.

What’s Next?

Generative AI offers transformative potential across industries by addressing unique challenges, streamlining processes and enhancing customer experiences. In the examples above, we’ve shown how businesses can leverage Gen AI to create innovative, scalable, and tailored solutions for various verticals. Each demo demonstrates how AI can drive efficiency, improve decision-making, and enhance customer engagement while maintaining compliance, de-risking and operational effectiveness.

To learn more about how these demos can be tailored to your needs, contact us.