Inspiration

Did you know 25% of startups fail within their first year, and 80% meet their end in five? The translation of a small idea into a thriving enterprise seems infinitely daunting at first, especially without the right guidance towards finding market fit and product differentiation. This is why we integrated both business consulting and content creation into one creative suite; a collection of generative and agentic AI tools for generating slide pitch decks, video ads and business roadmaps. UpStart aims to optimize and sharpen the vision of new startup founders, accelerating their workflow by hundreds of times faster when growing their business.

What it does

Our project is a one-stop-shop for prospective startup founders, using a suite of generative AI models and agents to help build a business from the ground up – all from a single idea.

Here is how it works:

  1. The User inputs their startup idea.
  2. Gemini 2.5 Pro takes the idea and generates a complete, 10+ page investor-grade presentation in Google Slides, complete with images and styling.
  3. The user's idea is converted into a script for a 8-second ad using Gemini 2.5 flash. Then, Google's Veo 3.1 model uses this prompt to generate a broadcast-quality video ad.
  4. The web app uses an AI-powered search (Gemini with Google Search grounding) to find a real, targeted list of potential VC firms or angel investors whose portfolios match the user's idea.
  5. The chat interface on the right-hand panel can be used to communicate with Everest, our custom fine-tune of Gemini 2.5 flash that has specialized knowledge/skills in advising startups. Everest is able to write and send emails to potential investors and aid the user with their startup journey.
  6. Everest is also able to create a specialized roadmap that is tailored to each startup to help founders create a solid business plan that significantly increases their success rate.

How we built it

We began by implementing the content generation, writing Python scripts to generate Google Slides, video advertisements and roadmap documents connected to their pitch. Each feature was powered by a specific API that connects back to Google Cloud, served to frontend through a Flask API.

  • Google Slides pitch decks were built by connecting the user prompt and context media initially to Google Gemini 2.5 Flash, which generates a JSON file for the Google Slides API to transform into a professional pitch deck.
  • Google Veo 3.1-powered video ads are generated with human product placement at the forefront, reaching target audiences through moving characters and cohesive text in a digital environment.
  • Fine-tuned Google Gemini model roadmap, providing a PDF of advice and time-bound objectives to expand your startup and penetrate the market
  • Everest: the ambitious, Google Vertex AI fine-tuned chatbot companion who is your loyal advisor to scaling up your startup business.

Visual identity was important as well: we built a sleek React + Vite frontend with glassmorphic panel displays, keeping users engaged as they upload a summary of their startup ideas along with any concept images or PDFs.

Challenges we ran into

Our biggest challenge was definitely fine tuning Gemini 2.5 Flash on Vertex AI to specialize in startup and business advice. We synthesized business and startup advice datasets in the JSON Lines format. The generated startup materials provide active feedback for the founder. System prompt engineering/tuning was also used extensively in this project. Due to us having many different types of models, varying from their complexity and usage, we had to make sure that each prompt fit specifically to each LLM and media generators in order to get the specifics we wanted. A lot of the times, the LLM would not listen to our request, so we had to come up with creative and unique solutions in order to curate the perfect prompt for the AI.

Also, a challenge we overcame within our team was the usage of code version control using Git. Since Git was relatively new to our collective experience – where we, at first, constantly modified the code at once – it was difficult to resolve the many merge conflicts that spread throughout the code. However, this in turn compelled us to communicate strongly with each other as we discussed which lines to keep and which ones to reject. This taught us not only how to use Git, but also how to work together in a tight-knit group: remaining patient with each other, even when it means resolving mistakes on another section of code.

Accomplishments that we're proud of

We were proud to effectively manage the time constraints of fine-tuning models using Google's Vertex AI, fitting its hours-long processes within our project timeline and integrating features effectively. We were also proud to figure out, after multiple software-focused attempt at a fix, that a CORS error was caused by port conflicts involving Mac system operations. This ensured our API remained operational across all computer platforms, while emphasizing how important it is to remain conscious of hardware.

What we learned

We learned how to fine-tune machine learning models. We learned how to build a full stack application from scratch. We learned to use Google AI Studio and Vertex AI to create useful media.

What's next for UpStart

We plan on making our existing AI functions more flexible and interactive for our users. For example, one idea we are experimenting with is letting users generate or upload an image and then use that image to create a video advertisement instead of relying solely on text. From there, we also plan to add the option for users to generate ads of different lengths.

Above all, a compelling idea we have is to generate a consistent visual theme that persists throughout all slides, videos and roadmaps. Using logos and color palettes, we can give founders extra control over their brand identity as they use our diverse toolset.

Built With

Share this project:

Updates