Building a Maker Mindset with AI

Hard truth time: AI is not a partner, coworker, or replacement. I hate seeing articles describing it as such. It’s a tool, pure and simple. Admittedly a transformative tool that can automate tasks, enhance decision-making with data, and amplify human potential.? But lets not get carried away.

We need to move past the buzzwords and recognize that most AI conversations today revolve around generative AI, particularly large language models that assist in coding, writing, contract reviews, and decision-making. But the real conversation isn’t just about AI—it’s about the people using it. Despite the hype and fear of AI replacing jobs, we inherently understand that high-risk decisions still require human oversight. You wouldn’t hand your pension over to an AI trading bot, nor would you trust an AI doctor without human validation. AI’s real value lies in augmenting human expertise, not replacing it—it’s a powerful assistant, but not the decision-maker.

Many businesses are rushing into AI adoption without fully understanding its capabilities and limitations, and the result is either an overreliance on AI or inefficient implementation. The corporate rush to delegate key responsibilities often stems from a lack of leadership comprehension. Leadership isn’t about delegation alone—it’s about knowing the craft. That’s why I stay hands-on, continuously learning and adapting to new tech like AI.?

I was going to write something just around this but it was getting a bit preachy and one thing I’ve learned since the advent of these generative tools is that “talk is cheap” - so instead I’m going to talk a little bit about something I’m currently doing and why.

So - time to put my money where my mouth is

Right now, I’m building an AI-powered meeting analyzer. I’m doing this because I believe AI should be a springboard—freeing teams from low-level tasks so they can focus on higher-impact work. There are other tools like this out there (Fathom springs to mind) but I wanted something that would solve some specific needs such as a persistent queryable knowledge base as an output. I also wanted to reinforce my “Maker Mindset” of building value, and getting a solid fundamental understanding of the tools we are playing with. The journey is as valuable as the destination as after all the ultimate goal isn’t to shrink teams but to maximize their potential. AI doesn’t replace leadership—it demands better leadership and better leaders understand the paths available to them.

Here’s a summary of what I’ve built so far. It's not very pretty, I’m sure there are better ways to achieve the same result but it works really well and I’ve learned a lot on the way.

Front End Interface

The front end is simple HTML sitting on a flask web server, using javascript for dynamic page updates. When it transcribes a meeting it makes a new folder based on the time stamp, and so displays these on the left in a simple list. The main transcription is shown on the main part of the screen, updating in real-time.?

Audio Recording and Transcription

Recording makes an audio file via FFMpeg - a powerful multimedia framework that supports real time audio processing. I then transcribe it from a staging audio file via Whisper. Whisper is OpenAI’s model to convert speech to text. It's very accurate, and is pretty good apparently for different languages, though I’ve only tested it in English. It does a good job so far even when multiple voices are involved but I’m looking at some speaker diarization and real-time alternatives as well for the future.

Once I have the text file, I then clean it up using Natural Language Toolkit (NLTK) and remove filler words to streamline it a bit (removing all the “basically”, “I guess”, “actually”, “literally” etc to reduce the size of the text without losing meaning)

LLM Analysis

At this point I send the transcription to openai (or GROK - does a similar job for this) and it summarises the meeting for me. It also identifies the attendees automatically (though it relies on name usage in the meeting so I’m looking at ways to pull this from calendar information) and all explicit and implicit decisions and action points.

Data handling

OpenAI is set to return the response in a JSON format. I like JSON as it's very readable, and easy to process for storage and queries. So I store the JSON file locally and then transform it via the front end to show the summary on the page. The reason I’m doing this is whilst this gets me to a rudimentary solution, I want to then do more with this. So the next step is to store the JSON file in Weaviate. Weaviate is an open-source vector database that's designed for AI-based semantic search. It looks for meaning rather than just exact keyword matches and it integrates really easily with Openai.?

Querying the database

Now I have my meeting data stored and easily accessible, I can then use the LLM to query it - for example “Tell me what we decided with Tier 1 competitors” or “Who was in the meeting last Tuesday?” or “Who did we agree to update the competitor analysis sheets?”

Meeting notes are no longer something stored on a server, never to be read again - they become something that is continually useful and part of a growing knowledge base to support the team and the account.

Why are you building this when tools like Otter and Fathom exist?

The key thing to this is the endpoint - the data-based storage in a way that is designed for continual updates and querying. I wanted that ability and something I could fully control, in a local environment. I didn’t want to be passing data to an LLM like ChatGPT which is designed to store messages (though I am going to look at something like langchain to store a contextual message history outside of Openai). OpenAI API does not store messages sent in this way - it remains stateless - so gives me some reassurance of privacy.

The other main and more important reason, however, is that of my own learning journey. I talk a lot about the intersection of Artificial and Human Intelligence because I am interested in the similarities of how we are building artificial structures of intelligence based on what we know about how the human brain functions. However, I wanted to go a bit deeper than this and get a better technical understanding of how real-world AI solutions are built. It's all well and good to have opinions but if you’re just using ChatGPT then you’re missing some amazing advances that are going on and it's extremely accessible.

It's also worth mentioning that building this simple tool above was done largely using natural language description. That means in simple terms I told ChatGPT what I wanted to do and it helped me figure out how to do it. Whilst I do have a background in programming, I’m antiquated these days With ChatGPT I’ve found a way to get rapidly up to speed and use a modicum of common sense, and the ability to read code generally as my secret weapon, rather than in-depth knowledge of specific languages. Maybe 80% of the work though was done for me - but like most generative AI, the remaining 20% is what will make or break it. I needed to understand every line of code, and I needed to be able to filter out when the LLM hallucinated, added unnecessary complexity or just took me down odd rabbit holes. Still - relying mainly on common sense and an infinitely patient AI, this took me 10 hours to get to a functioning useful tool that is already adding value.?That just blows my mind (and fills the lego-fan in me with joy)

All comments, suggestions, and criticisms about my terrible product design decisions along the way are welcome! What would you build? What's stopping you?

?