From Zero To SaaS #41 AI SaaS Integration

People who learn AI are amazed by its potential, while those who don't are bombarded with guilt-tripping content.

As an AI content creator, I can tell you that it is never as good as you think, but it is never as bad as people say.

Integrating AI into your business has to be a calculated move, like deciding who to marry, because you don't want to have a divorce later.

Although increasing profit, decreasing cost, and accelerating SOPs are good sales language, we should always build the product and the service around the user's need.

Take, for example, Apple Intelligence.

This app cannot even make an illustration, this example where you can spend 30b in R&D and still have product that sucks.

Never follow the masses unthinkingly.

UX of AI

When I offer an AI solution, I start with three personalised images to show the level of AI help.

Defined Input + AI Magic + Final Adjustment

✅ User doesn't have to think but can adjust the AI changes before downloading or sharing the end product.

⛔ Requires a great UI because it can easily overload the users with details it can adjust.

🚀 In Tella , you use their app to record yourself and your screen, click AI-generated zoom effect, and you better video. Afterwards, you can add, remove, or increase/ decrease the zooming effect.

Instruction + AI Magic

✅ The user must understand prompting techniques like negative prompting to generate data. This allows creativity and strategic people to transcend their standard outputs.

⛔ Needs much explanation and community to make the ball roll.

🚀 Midjourney , with a prompt, you can make any image you can conjure up, with many additional features.

AI Done For You

✅ The user gives input, and everything is generated, and you can't change details afterwards, but you can do a second call.

⛔ Very hard to do because you need to consider everything the user potentially wants even though it was explicitly asked for.

🚀 NotebookLLM, only a company as Google, can do this, you give it youtube link, text, any file, and it summarizes it for you. Although you can ask questions later, you cannot change the summarization process.

Currently building with Nick Broekema, Content Design SaaS tool sign up waiting list HERE.

LLM Calls

LLMs differ significantly from other machine learning models because the input and output use different methods.

Both use tokenization, where each word is represented by a numerical value to be sent to the neural net.

With the input, the model only needs to focus on how the word relates to its next of kin to extract the meaning if it can use parallel processing.

Like humans, it reads context and doesn't take the time to understand each word.

For the output, the model computes every possible direction, word for word, to select the most preferred output, which requires much more computing power.

Like painting, every decision, such as which color, element, style, etc., matters to creating an experience.

LLM Choice

Development

I use Gemini API unless they ask for OpenAI API for the applications I develop for my clients.

Being able to call model for free up to a hundred users per month is very advantageous because unless you do prompt engineering for three years, normal people can't notice the subtle difference between Claude and ChatGPT.

Agents

However, for agentic workforces on Relevance AI, I always use GPT4o for the subagents and GPT o1 for managing agents.

Other agent builders, such as Joep Baks , have already shown that GPT o3 mini, specialised in advanced reasoning, decreases agentic performance.

Security

If you are afraid of someone using your data, the preferred options in finance, mobile, health, etc, are the Azure OpenAI solution.

Microsoft stores your data on a secure cloud. It would be better to run your model on your local server, but you would need an NVIDEA cloud processor chip, which is expensive even if you rent one.

I am looking for companies with complex use cases they want to solve with agentic automation. Instead of conducting all these discovery calls with potential builders,

In this hackathon the participants compete for the contract. DM me if you are interested.