Demo overview

We conducted test marketing using NVIDIA's Nemotron-Personas-Japan to investigate "the reactions of financial institution customers to AI asset management advisory services." Nemotron-Personas-Japan is an open dataset for AI released by NVIDIA, containing a large amount of virtual human data from Japan, reflecting demographics, geographical distribution, and diverse occupations and attributes. By defining personas based on gender, age, region, occupation, and number, and setting the content of the survey, the survey can be executed, and evaluations and scores for the personas can be output. It is also possible to extract separate personas (e.g., young company employee novice investor / individual investor on the verge of retirement) and have the system explain "NISA and investment trusts suited to each individual." A key point is that, because it utilizes open data, "text can be tailored to the diverse customer profiles in Japan."

The flow of the demonstration

This demonstration uses personas to research and visualize reactions to financial products and services. It will proceed in the following five steps.

Step 1: Narrow down the persona

We create a survey target based on attribute conditions such as age and occupation.

You can adjust the survey population by gender, age, region, prefecture, occupation, and educational background.

Financial profiles are also automatically generated based on persona data.

The target audience is switched according to the industry and product concept of the target group. This makes it easier to demonstrate how changing the customer profile affects the results.

Step 2: Enter your research topic

We will develop proposals for financial products and services into questions that address interest levels, barriers to adoption, and multiple other factors.

By specifying a research theme and pressing the question generation button, you can quickly generate a list of questions that should be asked for this product proposal, based on factors such as level of interest, barriers to adoption, and priorities.

Question generation and survey responses are separated as model calls, allowing you to freely change the LLM model used.

Using thinking mode will lead to more accurate answers.

Step 3: View the answers live

The responses for each persona will be streamed sequentially.

Results can be observed at the persona level and question level, and the intermediate steps can also be shown as an experience.

You can check the overall progress of the survey, the score for each question, and the average score for each persona.

We'll transform waiting time into observation time.

Step 4: Read the trends in the report

We will compile the survey results by creating a distribution matrix of interest levels and barriers to adoption, selecting keywords, and summarizing the responses.

The response results are compiled and organized into a summary, polarity, barriers, and follow-up targets.

You can select key personas and translate them into points for discussion in subsequent business negotiations or follow-up interviews. The information is presented on a single, easy-to-use page and can be exported as JSON.

The results are recorded in the investigation history on the left and can be used to guide subsequent discussions. Since everything is saved locally, there is no need to worry about data leaks.

Step 5: Ask for clarification individually

I will continue asking follow-up questions about the personas that caught my attention in the report.

In addition to asking questions about points you are interested in, you can also have questions automatically generated based on your answer history.

You can get detailed answers based on the individual's experiences and preferences, tailored to their persona.

All question and answer history is saved locally.

Hardware used, etc.

- UI: Open the demo screen in a web browser
GPU server: NVIDIA H100 GPU x 1 Or NVIDIA DGX Spark™
• Inference engine: vLLM
• Model used: NVIDIA-Nemotron-Nano-9B-v2-Japanese
Persona data: Nemotron Personas Japan
・ History saved: Blackite

structure

This is a combination of a research app, persona data, and a Japanese LLM (Language Level Master).

When a user enters a research topic on the screen, the app asks the same question to multiple personas.

The AI will respond by fully embodying each persona.

The answers will be displayed on the screen in order, and finally, the overall trends will be summarized in a report.

Furthermore, you can select a persona that interests you and ask additional questions.

This system involves a human selecting survey participants, AI simulating their responses, and then compiling the results into a report. It's a demonstration for the financial sector that covers the entire survey process, not just a single chat.