Pitch Arena

Inspiration

Pitch Arena came out of the general experience we have with many LLMs: you enter an idea and lo and behold it loves it:

• wow thats an amazing idea
• love it would you like to vibe code the MVP?
• you should quit your day job and do that right now

Ok, the last one I made up but the problem of getting such positive feedback yet the hard reality check when we present can be a challenge to reconcile. Pitch Arena tries to capture that hard 'shark tank' like realism that you get when you present an idea in front of a bunch of dubious VCs, hackathon judges or grumpy friends and colleagues.

The name Pitch Arena describes the idea of having to present in front of real people with agendas and objectives; they are not there just to confirm how wonderful you are. You enter the arena in order to 'fight' and win because whether its in front of VCs, a jury or if you have ever been in a booth you know you have to convince real people.

What it does

Roughly: Pitch Arena uses AI characters that are 'realistic' meaning they have personalities, objectives and a tone or mood. They are there to question you in a hard fashion in order to get a more realistic pitch review. The arena itself is where you test your idea out and the judges are there to verify that your idea matches to the arenas agenda. (For example in the video we can see me trying to get pitch arena to pass the gemini hackathon criteria.)

Starting from the top:

• you create an arena which is a configuration containing goals/objectives, constraints and judges
• judges have their own tone and areas of interest and they are there to assess whether you are aligned with the overall goal but from their perspective. A CFO won't ask the same questions as a CTO but they are both aligned on the overall agenda.
• users enter an arena and can either upload an initial pitch e.g. the project overview, slides etc or they can enter direct and explain everything from scratch.
• interaction for the moment is though text or voice (TTS and STT)
• once its all over the user gets a scorecard that contains two parts: the judges assessment of the idea and a second coach assesses the quality of the speech itself.

The goal is to give you an 'unsafe' space to test your ability to pitch and answer hard Q&A. So we assess the idea in relation to an objective (win a hackathon) and how you presented it (the pitch coach) as its not enough to have an idea: you need to convince too.

How we built it

The application was built using Firebase and Angular:

• Angular UI deployed to Firebase
• Gemini via firebase SDK
• elevenlabs for the TTS and STT (via functions)

The interaction with the arena is via a voice/text interface for the moment. A multi pass prompt handles the context where each judge has a set of criteria and updates those criteria given a users answer to questions.

Example: idea has a large market of paying customers

The judge will ask questions on this topic till it feels that the criteria is satisfied. The judge can be tuned in order to avoid bullying but it can mean the criteria are unfulfilled.

At the end of the arena (round based for the moment) the judges create a scorecard saying their criteria were met/unmet and their overall confidence. The pitch coach will also score but is there to evaluate the users general presentation and answering skills.

Challenges we ran into

The hardest part by far is creating the prompts and managing the flow in order to get a responsive feeling We iterated several variations to try and balance the intelligence/realism of the room against performance. If the arena takes +10 secs a question it becomes unrealistic (and boring).

• We use a mix of parallel questioning and host interjections to try and get a smoother feeling.
• Also using Geminis thinking budget helps a lot

Accomplishments that we're proud of

Its strange but when I challenged Pitch Arena in Pitch Arena for the first time I got annoyed with the jury and started typing in a lot of long winded details ... I then realised it was working as I was engaged and needed to calm down ;-)

What we learned

Prompts with multi state conversations are key to getting something working and its really there that I learned a lot. The notion of a cross round score card and ensuring questions aren't repeated was valuable for nearly all complex conversational models.

What's next for Pitch Arena

I think that there are a lot of directions that can be taken:

• the concept could be interesting for incubators, VCs, accelerators in order to screen candidates
• using more of gemini by adding in RAG and/or search in order to pull in information for overall idea assessment (for example allowing the bots to assess the market situation)
• integrating into a video conferencing system with perhaps realtime characters

Also it doesn't have to be for pitches as the configuration system is quite open so it could be using for training, interviews, sales teams ... anywhere where you want to test being challenged.

Built With

• angular.js
• elevenlabs
• firebase
• gemini