Bronco

User enters Instagram username.
Gets recommendations for their essence.
Diagram of flow

Inspiration

Rohan, Matthew, and Jonathan grew up in the same town in Illinois. We came with an interest in AI agents and enhanced personalization in software. We spoke with several sponsors and identified key problems in industry that can now be solved for the first time using AI agents.

We learned from the COO of Apify that businesses use Apify to pull user context to provide personalized experiences. Private equity firms need additional agents to enrich this data for due diligence.

We learned from e-commerce store owners and software applications that businesses and applications face a "cold-start" with no data on new users and are forced to provide generic experiences.

What it does

Our software solves the cold start problem by:

User enters site hosted by Render
Apify actor ingests user social context
Gemini backend research AI Agent analyzes user context to determine user preferences for unique application (i.e. food preferences, dating preferences.)
Agent self-learns and improves prompts based on LaunchDarkly performance analyses
Through testing, prompts are autonomously improved by an agent that judges responses and user activity (clicks, conversions, etc).
Site bootstraps a recommendation system from scratch.

How we built it

We used a frontend that intakes username. The username passed into our service, which then returns a list of recommendations. On the front end, it will send the user preferences back to our service, which would then pass it into a judge model, which would pick out what went wrong with the prompt, and readjust the prompt, and use that prompt for future users. It's continuous feedback for new users to the applications.

Challenges we ran into

Similar to smaller e-commerce stores, we do not have extensive data sets that large companies use to optimize recommendation systems.

In order to surpass this hurdle, we bootstrapped an AI agent that autonomously self-improves recommendations by judging recommendation quality and evolving the prompts to provide enhanced recommendations for any site.

Accomplishments that we're proud of

Many e-commerce website rely on pre-trained models that rely on on-site clicks and shallow behaviors. Our agents successfully leap over these hurdles and provide a step-function in personalization.

What we learned

We were really impressed with Apify. We were able to gather context from a variety of sources. We stayed narrow in our initial features but can see ourselves expanding context broadly.

What's next for Bronco

We are going to seek customers who may want to eliminate high-friction onboarding or shallow personalization to immediately provide intimate personalization at scale.

We also want to implement Sherlock, which we learned is a valuable tool that, given a username, can detect similar social profiles and aggregate context into one pull for user analysis.