Me 2.0

Inspiration

The goal was to explore way to scale yourself – beyond just automation, but scaling your expertise + specific knowledge (your unique value).

What it does

Me 2.0 is a digital cloning platform that lets you create a clone that acts like you, has your knowledge, and interacts w/ others on your behalf.

We targeted one niche w/ one specific use-case to get started: online coaches + people who spend their time helping people achieve a specific outcome based on their own unique insight / knowledge.

A coach can [1] clone their likeness + knowledge based on the content they've already created, [2] fine-tune how their clone answers by talking to it + correcting answers w/ natural speech, and [3] letting their clients subscribe to their clones + interact w/ them 24/7 whenever they want.

How we built it

We had a simple flow:

1) clone knowledge – we did this manually (for now) by transcribing podcasts of a relationship coach, chunking + embedded the data into pinecone, and setting up retrieval w/ chatGPT

2) setting up chatGPT to respond w/ the "likeness" of the coach (doing some interesting things like asking chatGPT to analyze the transcripts + engineering / iterating on prompts)

3) voice cloning (using 11labs)

4) setting up + hosting a simple front-end for speech w/ voocode + replit

Challenges we ran into

Honestly, hacking this together was hard lol.

Trying to automate chunking, embedding, and retrieving data w/ pinecone + chatGPT was a challenge. Even then, it was hard to measure how effective adding the external DB was in improving our responses.

ex. How can you tell if the response based on a persons content is better than what chatGPT would give you out of the box? What even is better?_

These questions are pretty subjective + hard to answer, but overall we bet people would want to subscribe to these clones b/c they resonate w/ the unique perspective of the creator themselves – the same reason why we choose to follow one influencer / content creator over another doing similar things. We're fans of people by nature, and we enjoy the creativity + uniqueness of individuals – not just getting answers from a chatbot.

Getting the realtime voice to work back and worth was also annoying af to debug. There has to be a better package out there to get this to work smh.

Accomplishments that we're proud of

Honestly, we're stoked that we got GPT to respond like Jessica + her knowledge – a relationship coach w/ lots of content out on the internet, but not famous enough for GPT to understand how to mimic out of the box.

We did it once, so how we have a playbook to do it w/ anyone + a pretty simple roadmap to automate the whole process – that could be gamechanging.

What we learned

From a product level, we understood the limitations of using LLMs to respond in someone's own likeness + from their perspective. It's easy to read about on twitter, but actually playing w/ the prompts + responses teaches you a lot about how GPT actually learns.

Fine-tuning LLMs may add another layer, but it's a tradeoff b/w time + quality.

From a personal level, we learned how insanely cool it is to build alongside some wicked product builders. When we ran into an issue all we had to do was ask someone around us – chances were they faced something similar before, or knew someone who had. 100x learning moment.