How to Use GPT-Based AI-Agents as Founder or Investor: The "LeanStartupAgent"
DDVC #36: Where venture capital and data intersect. Every week.
👋 Hi, I’m Andre and welcome to my weekly newsletter, Data-driven VC. Every Thursday I cover hands-on insights into data-driven innovation in venture capital and connect the dots between the latest research, reviews of novel tools and datasets, deep dives into various VC tech stacks, interviews with experts and the implications for all stakeholders. Follow along to understand how data-driven approaches change the game, why it matters, and what it means for you.
Current subscribers: 9,414+, +746 since last week
Brought to you by VESTBERRY - A tool designed specifically for data-driven VCs
Watch our 7-minute product demo, showcasing the platform's powerful features and intuitive interface. Gain valuable insights and make data-driven decisions with unparalleled ease.
I’m incredibly excited to welcome more than a thousand new readers who joined since launching the “Data-driven VC Landscape 2023” last week 👋🏻 Of course, I always try to understand and tailor my content as well as possible to you, the community.
Following the recent hype around GPT-based agent experiments like BabyAGI or AutoGPT (the hype within the hype🤔), I’m excited to have Professor Dries Faems contribute today’s guest post on how to build and use your own AI agent with Langchain showcased by “LeanStartupAgent”. Of course, the concept can be applied to any other use case too!
Some context: Dries is a Professor for Entrepreneurship, Innovation and Technological Transformation at the WHU Otto Beisheim School of Management, one of the leading entrepreneurial universities in Europe that is lucky to count the founders of Zalando, Rocket Internet, Forto, Flixbus, HelloFresh, and many more unicorns to its alumni.
Thank you, Dries, for sharing your innovative work with us and providing a blueprint in your guest post below 🙏🏻
The “LeanStartupAgent”
In this article, I will introduce LeanStartupAgent, which is a GPT-based application that allows to automatically generate business models for founders. First, I will briefly summarize the theoretical foundations of this agent. Next, I will elaborate on how this agent leverages novel opportunities in natural language programming to automate a wide variety of data-driven business activities. Finally, I will give a clear instruction list on how you can use the agent yourself.
Theoretical foundations of the LeanStartupAgent
LeanStartupAgent is an application that helps entrepreneurs and innovators to automate several steps in business model development. The logic of this agent heavily relies on the Lean Startup approach, which has been popularized by influential scholars and practitioners such as Eric Ries, Steve Blank, and Alex Osterwalder. The lean startup approach is a methodology for developing businesses and products with a focus on efficiency, customer-centricity, and continuous learning. Here are the core principles of the lean startup approach:
Identifying Customer Pain Points: The first principle of the lean startup is to identify the problems or pain points that customers are facing. By deeply understanding these pain points, entrepreneurs can gain insights into unmet needs and opportunities for innovation.
Failing Quickly through Hypothesis Testing: The lean startup encourages entrepreneurs to formulate hypotheses about their business model, product features, or target market and test them rapidly. Instead of spending extensive time and resources on developing a full-scale product or service, lean startups aim to create minimum viable products (MVPs) to validate their assumptions and gather feedback from real customers. This process allows early detection of flaws or misconceptions, enabling quick adjustments and avoiding significant failures down the line.
Clear Value Proposition: A value proposition is a clear statement that explains how a product or service solves a customer's problem or meets their needs in a unique and compelling way. In the lean startup approach, it is essential to formulate a well-defined value proposition that resonates with the target audience. By analyzing the pain points and understanding how the value proposition directly addresses those challenges, entrepreneurs can communicate the benefits of their offering effectively. It is crucial to evaluate and analyze the fit between the value proposition and the customer's pain points. The value proposition canvas is a structured tool to formulate and evaluate value proposition-market fit.
If you want to get a deeper understanding of the core principles, check out the podcast episodes with Steve Blank on ‘Decoding innovation theater’ and Alex Osterwalder on ‘The art of hypothesis testing in startups’ in the Most Awesome Founder Podcast.
Technological Foundation of the LeanStartupAgent
The LeanStartupAgent uses prompt templates, chains, and agents from the Langchain Python library to generate business models based on user input. Below, I explain these concepts in more detail:
A prompt template is a string that contains placeholders for variables that can be filled in with different values. For example, "What is a good name for a company that makes {product}?" is a prompt template that can be formatted with different products. Below you see the code of the first prompt template in the LeanStartupAgent, which uses the input of the user (i.e. job to be done and customer description) to generate a description about the core pain point (i.e. hardest part of job-to-be-done for customer segment).
A prompt chain is a sequence of prompts that are connected by logic and context. For example, a prompt chain can ask the user to define their customer segment, then their value proposition, then their revenue stream, etc. In the LeanStartupAgent, we build a chain where the pain point is used to automatically generate a value proposition. This value proposition is subsequently used as input to generate the value proposition canvas and to identify core competitors.
An agent is a stateless wrapper around a prompt chain that takes care of formatting the prompts and parsing the responses from the chat model. For example, an agent can take the user input "I want to make an app that helps people learn languages" and format it into a prompt chain that asks the user to specify their target market, their unique selling point, their pricing strategy, etc. Langchain also has dedicated tools that agents can use to execute particular tasks. For instance, you can define a search tool that allows an agent to execute particular search tasks on the internet. In this way, you can start combining formulating prompts with specific tasks. For an overview of existing tools that agents can use, see here. In the LeanStartupAgent, we use the search tool of the SerpAPIWrapper to generate additional information on the companies that were identified in the prior step.
Some additional comments for the coding enthusiasts among us:
It is not necessary to the use the OpenAI LLM to use Langchain elements. It is also possible to use open source LLMs. However, my experience is that using the OpenAI LLM seems to be the best option when using agent tools.
Hugging Face just announced the launch of its own Transformer Agents. They are likely to become a huge competitor for Langchain.
How does LeanStartupAgent work
LeanStartupAgent uses the described Langchain components to create a dynamic and interactive application that guides the user through the process of creating a business model canvas.
Starting point:
You need to access the code in the following github repository: https://github.com/DriesFaems/LeanStartUp-GPT-Agent. You need to download the file LeanStartup GPT Agent public.ipynb and subsequently you can run it in your own python program.
You need an OpenAI API Key and a Serp API Key to execute the code
You need to input your job-to-be-done and your target customer segment in the code
Output generated:
The code creates a text file, containing the following information:
Hardest part of job-to-be-done
Value proposition
Value proposition canvas
Three potential competitors
Additional information on the competitors
Using Streamlit, I also created a webapp version of the LeanStartupAgent. To run this version, you need to use this Collab notebook. You can see a demo of the app via this YouTube link:
Conclusion
Prompt templates, chains and agent tools are creating fantastic opportunities to leverage large language models in ways that we could anticipate some months ago. With some limited coding capabilities, you can now create and automate sequences of tasks for very specific circumstances. The application opportunities in the domain building and analyzing startups seem to be unlimited with this kind of approaches!
On May 25, 2023, at 8:00 P.M. GMT+1, I will hold a free online seminar via Zoom on how ChatGPT can be utilized for a wide variety of business cases such as customer exploration, negotiations and competitor analyses. Those interested can sign up here.
This is it for today. I hope you enjoyed this more practical episode and got some ideas on how to develop your own agent.
Stay driven,
Andre
Thank you for reading. If you liked it, share it with your friends, colleagues and everyone interested in data-driven innovation. Subscribe below and follow me on LinkedIn or Twitter to never miss data-driven VC updates again.
What do you think about my weekly Newsletter? Love it | It's great | Good | Okay-ish | Stop it
If you have any suggestions, want me to feature an article, research, your tech stack or list a job, hit me up! I would love to include it in my next edition😎