PROPL

Propel Your Product Pitch Forward with our AI Testing Platform!

Comment

Inspiration

Every company needs to do some form of marketing for people to know about their products and services. However, attempting to market a product without an understanding of what demographic is best served by the product, and what is the most effective method to entice them to a product, will very likely lead to failure, incurring thousand, if not millions of dollars lost for a company. However, market analysis can be costly and time consuming, with a single high quality analysis project taking hundreds of thousands of dollars and weeks of time.

This is where PROPL comes in. With the rise of multi-agent AI systems, we saw incredible potential in simulating demographics of people using LLM agents, and having AI agents simulate both buyers and sellers. This would allow for marketing campaigns to be simulated and analyzed, allowing businesses and marketers to gain insights about their ideas in minutes, and not weeks.

What it does

PROPL is a product and marketing testing platform that uses the power of multimodal, agentic generative AI to simulate marketing products to many different types of people, and analyzing the effectiveness of selling the product to different demographics.

A user provides a product name, description, important highlights about the product, and any multimodal advertising material they wish to test. Once submitted, these are given to a "salesman" agent, who is tasked with attempting to sell the product to a variety of different customers, each with a unique persona. Multiple conversations are simulated in parallel with the buyer and customers, and once all the conversations are finished, a different agent analyzes each conversation for their success in selling the product, general conversation sentiment, and top reasonings provided by the customer.

These statistics are compiled together into demographics analytics, providing the user with charts and explanations on what groups of people were most interested in their product, and what was most and least effective about their existing pitch. Finally, the best performing conversation is also generating into a voiced conversation, letting users hear how their product is being pitched and how a customer may react.

How we built it

Image

We designed our project around two concepts: engaging and visually pleasing front end, and a powerful multimodal, multi agent back end. For the front end, we built it using React, as it gave us a lot of flexibility to design the visuals and interactivity of the website, as well as generating visually appealing plots. For the back end, we built our multi agent system off of Gemini 2 Flash, with multiple instances of the LLM being initialized with different roles for salesman, customers, and analysis tools. Multiple copies of the salesman agent are sent out to each of the customers, which each have a randomly generated persona. The salesman is told out the product they are selling and the customer they are selling to, while the customer is told about their persona. All the conversations are simulated at once.

Once all the conversations are generated, we similarly call Gemini with the conversation as history, asking if the sale was successful, a sentiment score, and highlighted reasons on why the sale was successful or not. These results are then analyzed for demographics information and the best scored conversation is saved separately. The demographic analytics are returned to a separate website page to create graph visualizations to explain the sales results to the user.

The saved conversation is fed to a separate API endpoint that uses AWS' Polly text to speech system. Separate voices are given to the salesman and customer, matching the original customer persona. This is then also returned to the website, which syncs the avatars to the voice generation conversation to provide a visually clear conversation between the salesman and customer.

Challenges we ran into

We ran into an initial challenge at the start in what our input would be. We originally wanted to input startup pitch decks to find the most similar pitch decks from successful startups, and offer improvements to your own pitch decks. However, this was not easily feasible, as pitch decks are heavily reliant on visuals, and two pitch decks could look very visually different but contain similar concepts. Therefore, we switched to a product pitch, with multimodal advertising material as a supplement.

We struggled with setting up the multimodal input into the agent system, as Google's API SDK documentation did not provide a clear explanation on how to do it. We also had a significant roadblock with connecting the website's graphs to the agent API results, which broke during development and took quite a bit of time to fix.

Finally, we had difficulty setting up the text to speech and avatar system, as we had to adjust our presentation based on the cost of APIs we originally wanted to use, and had to look for alternatives, such as AWS's Polly Text-to-Speech.

Accomplishments that we're proud of

We are very proud of the visually appealing website, complex multi agent system, and voiced conversation. We were committed to not only providing a quality service, but one that the user wants to keep using, which a user-friendly and appealing website greatly contributes to.

This is also the first time we built a multi agent system at this scale, generating this many conversations at scale, and analyzing the conversations together to provide useful analytics to users. We spent a lot of time brainstorming how we could implement so many agents interacting together at once, and are satisfied with the results.

Finally, the voiced conversation is a unique feature we all love. The voices provide a "real life" example of the product, showing off the qualities of the product beyond statistical metrics. We spent a lot of time searching for the best way to create natural sounding recordings, and found AWS' Polly API worked really well.

What we learned

For this project, we spent much time confirming our knowledge on developing front end and API frameworks, and learning how too extend it to a larger scale. We really had to learn about how to use mutli agent AI at a large scale, using up to potentially hundreds of concurrent agent conversations at once efficiently.

We also learned a lot about what makes a good product pitch, and what is most important to customers. We found that customers tend to be very concerned when high prices are brought up for products, and that many conversations dealt with convincing customers that the high price was worth the quality and usability of a product.

What's next for PROPL

We see a lot of potential with taking PROPL beyond a hackathon project, and PROPLing to an actual future startup. We believe PROPL is a unique and incredible offering that can impact a highly important, high capital niche.

Our current offerings allow users to write their product pitches and provide pre-made advertising material alongside to test their viability across different markets and provide analytical insights on their pitch's impact and areas of improvement. For our next steps, we want to take this beyond just insights and suggestions for improvement, and generate the improvements ourselves.

We envision PROPL becoming a platform for testing and automatically improving pitch concepts and advertising. An immediate potential addition to PROPL include an iterative pitch testing and improvement loop, where automated analysis is fed back to the pitch, tweaking it to improve found deficiencies in the previous iteration. Users could then watch how the pitches improve across analytics metrics until they meet a desired threshold. A second addition is to generate new advertising material, including images, videos, and campaign ideas. This would elevate PROPL beyond just an analytics and pitch tool, but provide users an entire multimedia ad campaign, proven to appeal to a wide demographic of potential customers.

We have a strong belief that PROPL can become the next tool for businesses and advertisers to create, test, and improve entirely new product campaigns, improving efficiency and exploding market potential.

Built With

Share this project:

Updates