It’s official — we just launched the world’s first reasoning AI developer. My goal with this post is to explain what that means and why we did it.

Databutton’s reasoning AI developer is the first to ever

Learn and know about your app
Share your app brief as a vision, specs, and design inspiration to get started. Databutton has a long-term memory of your app and decisions made along the way

Create high-level plans
Get an instant development plan, broken down into clear, actionable tasks. Your plan evolves as you keep giving feedback and refining ideas.

Execute big tasks
When you hit “start task” the AI agent will think, research, code, and only ask for input when it needs your guidance.

Take technical ownership
From deployment choices to database selection, the AI agent handles all the heavy lifting — while giving you the option to override any decisions whenever you want.

Why Databutton 2.0?

The first generation of AI coders have been incredibly powerful and strong enablers for builders around the world. For developers, they’ve increased speed substantially and for non-technical builders, they have helped bring ideas to life (and were used for a lot of prototyping, if we’re being honest).

At Databutton, we saw the opportunity to make a real dent in the ecosystem with an AI developer for non-technical folks to build SaaS or radically transform their businesses. In order to do it, we needed to rethink the working model and overcome today’s weaknesses.

What was Databutton 1.0?

We launched Databutton, the first agentic full-stack AI developer, in June 2024 after starting work on it in October 2023. Back then, code writing capabilities from LLMs were good enough for developers to use.

To use Databutton as a non-developer, you had to really want it. Through the experience of building, you became a quasi-developer. And we provided a lot of support to help people build absolutely incredible applications (examples 1, 2, 3, 4).

Personally, as a non-developer, I needed plenty of help from Martin, our CTO, to get across the finish line on apps I built. And then suddenly, Claude Sonnet 3.5 represented a huge upgrade in code generation quality. Non-technical people like me were able to generate working code consistently.

What we saw through that coding upgrade, though, is that there is still a huge gap between generating code and being a builder.

The problems with today’s AI coders

In order to benefit from an AI coder today, you have to be a builder. What does that mean?

• You spend a majority of your time building
• You need to keep track of your app and plan
• You need to make all the technical decisions

Meanwhile, the AI itself:

• Has limited knowledge about the app/goal

As a user, you start out with a prompt. But where is the brief? Or the spec? You and the AI coder are never really aligned on anything beyond the the prompts you send. A lack of alignment on both decisions made in the past and the eventual goal result in the need for reiteration all the time.

• Doesn’t work off an app-wide plan

If you’re non-technical and don’t have help creating a plan, you waste time on implementing the wrong features at the wrong time.

• Needs to be sent one prompt at a time

If you’ve used codegen software, you’ve definitely had the following experience. It takes quite some time to become proficient at using these things.

Reasoning — the gamechanger

We started Databutton to enable those with the expertise and problems to create the software they need.

However, if you’re non-technical and really want help with software, you often look for a tech partner or CTO. This is because you are focused on your day job! Your value add.

Our aim was to replicate this relationship to our fullest ability with the world’s first reasoning AI developer. Instead of commanding, you collaborate with Databutton’s AI agent.

Your role is to lead the project and give direction. What is the objective, what problems are we solving, and how should the software work? You come with user feedback and needs.

Databutton’s AI agent’s role is to come up with an execution plan and build the software, just like a technical partner would.

The timing couldn’t be more right — the advent of reasoning LLMs made it possible for us to change the game.

Scroll back up and check out the video a second time 🤓

Have thoughts on this direction and what’s coming next?

Try it out for free and let us know what you think!

Why we built the first reasoning AI developer was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

Agents are intelligent programs capable of achieving complex tasks by taking targeted actions.

These AI agents aren’t just useful — they can be insanely powerful.

They execute specific functions with impressive efficiency.

Agents are intelligent programs that achieve complex tasks by taking actions.

• But how simple is it to actually build and integrate them?
• And what’s the best approach to making these agents part of your full-stack applications?

This week, I spent most of my time in building AI agents from scratch and using them to create scalable, full-stack apps. Here’s what I discovered ↓

What I’m Using to Build Apps Fast

To speed up app development, I use Databutton. It does simplify and speeds up in building scalable, real apps with AI, making it a powerful AI App Builder (in my biased but honest opinion!)

Databutton’s backend magic comes from leveraging Python with FastAPI routers. This setup provides easy access to any Python libraries for generating APIs easily.

For the frontend, Databutton uses ReactJS enhanced with shadcn/ui for that extra visual flair.

So, to start building the app? It’s all about prompting in simple english.

The UI can be built here with a simple prompts (quick tip: pass an existing screenshot or images of the app in mind)

To build the backend quickly, I passed Databutton with phidata’s docs over the prompt ( passing code snippet works very well) .

With a little guidance, Databutton researched the relevant docs, extracted essential context, and generated the backend code efficiently.

How I’m Building AI Agents

For building agents, I’ve found Phidata’s Python package incredibly intuitive, thanks to its clean and straightforward syntax.

Phidata : An open-source platform to build, ship and monitor agentic systems.

Here’s a code snippet to build a simple YouTube Agent:

# Import Relevant modules
from phi.agent import Agent

# Agents having accessing to tools
from phi.tools.youtube_tools import YouTubeTools

# Initialize YouTube agent
agent = Agent(tools=[YouTubeTools()],
            show_tool_calls=False,
            description="You are a YouTube agent. Obtain the 
              captions of a YouTube video and answer questions. 
              Provide concise, focused answers and use markdown formatting for better readability."
        )
# Get agent's response
response = agent.run(prompt)

The Agent class from Phidata makes it super simple to customize agents with a variety of parameters.

You can read more here about the available parameters.

Additionally, Phidata comes with in-built tools that helps in achiveing multiple tasks .

Toolkits - Phidata

In Databutton, the agent wraps the Python function into a FastAPI router.

Once Databutton tests and ensures the backend runs smoothly, the next step is integrating it with the frontend using its internal module.

This allows to connect the backend APIs with the app’s user interface.

You can find the entire backend code that Databutton generated here.

When everything is ready, deploying the app takes just one click 🚀

If you’re curious, check out my recent YouTube videos where I walk through creating these agents and build the app from scracth.

Infact, with Phidata, you can even build teams of agents working collaboratively. I elaborated on this in one of my last livestreams where I built an app live.

Thanks for reading! If you found this helpful, subscribe to my newsletter for more insights and updates.

How to Build Full Stack Agentic Apps was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

Can’t read this? Use our friend link.

When Chintan first wrote to us on Intercom about his app, Privacai, we knew he was building something important for the world. What we didn’t realize is that he’s also Dr. Dave — a dual-certified Internal Medicine and Critical Care Physician passionate about sharing his learnings in medicine to the tune of nearly 60k TikTok followers.

His feed is filled with incredibly useful medical tips from his years practicing as a doctor. Here’s an example:

Consumer disempowerment

Corporations are introducing more and more far reaching clauses in their terms of service and consumers simply do not have the tools necessary to understand their own rights, or rights they are giving away. Having the time to read 30+ pages of legalese or feeling pressure to sign documents without having appropriate time are real issues people face every day. Informed decisions are hard to come by.

Some particularly upsetting news related to this exact topic has been making their rounds in the news lately.

Disney joins other companies such as Airbnb and Walmart that are using increasingly aggressive strategies…customers more and more must agree to contracts with sweeping arbitration clauses to use their services, but the consequences can be larger than they can be expected to comprehend.

“The Average Joe in society doesn’t know what arbitration is, let alone understand the content of what they’re signing,” said Creighton University Law Professor and arbitration expert Hossein Fazilatfar. (CNN)

Below are two examples of these “sweeping arbitration clauses” causing incredibly negative real world consequences for consumers.

• Disney+ terms prevent allergy death lawsuit, Disney says
• Couple in a severe Uber crash can't sue because of an Uber Eats order | CNN Business

Dr. Dave’s own experiences

Dr. Dave is both a provider and consumer himself, so he acutely feels these problems in his day to day life:

As a practicing doctor in the US, every procedure or treatment I initiate, we have to obtain informed consent from the patient. Informed consent means that the individual is aware of both the risks and benefits of all options available to them. Terms and conditions and various policies that we all agree to online are always too long and anyone rarely reads them. However, they have real world and negative implications.

Recently, I was at a car dealership to lease a car for the first time. It was close to 5pm so the agent kept hurrying to get all the paperwork sorted out. However, I was given 10+ pages to read and sign within 5 minutes that would lock me into a 3 year lease agreement. I just wanted to know if there were hidden fees or other important things to consider before signing it.

- Dr. Dave

The opportunity with PrivacAI

PrivacAI was born out of these personal experiences as well as a general trend of consumer disempowerment. As such, their mission is to empower individuals to make informed decisions about the policies they agree to. Large companies currently hold unilateral power in changing the policies whenever and however they want.

By providing awareness and accessible information to individuals, PrivacAI can collectively prevent disempowering and over-arching terms and conditions in policies that have become far too rampant currently.

How it works

First you land on Privacai.io — you can make informed decisions either based on looking up a company’s policies via search or by pasting in a link to any privacy policy, ToS, data policy, and more.

If you pick a company, you can choose to narrow in on a specific document or analyze the entire company’s policies based on your values. You can select from several pre-defined values or write about your own values by visiting the profile section.

Get a summary of how the company’s policies conflict with your values to make more informed decisions as a consumer!

Achieve your own mission

We’re incredibly proud to be Chintan’s service of choice for building his app and achieving PrivacAI’s mission. Need help with your own app or software? You can also sign up for Databutton or sign up for our App Launch Program!

Dr. Chintan Dave is on a mission to protect consumers from predatory policies was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

How to Build AI Apps with Multi-Agent Systems Using OpenAI Swarm Library 🐝

The world of multi-agent frameworks is expanding, and a new player has just joined the scene: Swarm.

Is it truly something new? Well, yes and no.

So, what is Swarm?

Swarm comes from the OpenAI Solutions team, and it’s designed to make working with multiple AI agents easy and intuitive.

An educational framework exploring ergonomic, lightweight multi-agent orchestration.

Definitely give this detailed cookbook a read!

Orchestrating Agents: Routines and Handoffs | OpenAI Cookbook

Swarm’s magic lies in simplifying the tricky process of using large language models for agentic behavior.

Think fewer large prompts, function calling, and less struggle.

When working with language models, quite often all you need for solid performance is a good prompt and the right tools. However, when dealing with many unique flows, things may get hairy.

Key Concepts behind Swarm

First up, Routines/Agents.

The idea of a “routine” in Swarm isn’t rigidly defined. It is more like a flexible set of steps.

As the OpenAI team puts it:

The notion of a “routine” is not strictly defined, and instead meant to capture the idea of a set of steps. Conretely, let’s define a routine to be a list of instructions in natural langauge (which we’ll represent with a system prompt), along with the tools necessary to complete them.

Another important concept is Hand-offs.

This is where one agent passes an active conversation to another agent, similar to being transferred during a phone call — except these agents know exactly what’s been discussed so far.

The cookbook describes it like this:

Let’s define a handoff as an agent (or routine) handing off an active conversation to another agent, much like when you get transfered to someone else on a phone call. Except in this case, the agents have complete knowledge of your prior conversation!

Here’s a simple way I thought about it, so I created a quick illustration 💭

The swarm Python package is simple yet powerful.

In a nutshell; you define a Swarm client, add agents, and run it.

It’s that straightforward!

Here’s a quick example code snippet to get started.

from swarm import Swarm

# Instantiating a Swarm client 
client = Swarm()

# Define Agents (Check the example below)
# Agent_one
# Agent_two

# Run Swarm
client.run()

Note: Swarm’s run()method orchestrates agents, manages hand-offs, and handles context—all in a way that makes working with multiple agents feel intuitive.

Defining agents is key in swarm .

# let's consider these two agents

english_agent = Agent(
    name="English Agent",
    instructions="You are an expert in English. Respond in English.",
)

spanish_agent = Agent(
    name="Spanish Agent",
    instructions="You are an expert in Spanish. Respond in Spanish.",
)

Next, handling “hand off” between / among the agents.

# Defining a function for handoff
def transfer_to_spanish_agent():
    return spanish_agent

# Appending functions to the agent
english_agent.functions.append(transfer_to_spanish_agent)

I’ve already built a multi-agent backend using swarm with a real use case in mind.

Check out the detailed walkthrough in this video and, I’ll also introduce the key idea in the next section below!

This what I built

In less than 30 minutes, I could build a multi-agent chatbot called HealthBot Assistant.

This chatbot uses different agents, each specialized in a specific task.

When a user asks a question, a Triage Agent (also we can refer this agent here as the master agent) determines which agent is best suited to assist them.

triage_agent = Agent(
    name="Triage Agent",
    instructions="You are a healthcare triage agent. Help users describe their symptoms and determine which healthcare-related agent they should speak to (e.g., for medical advice, scheduling an appointment, or checking prescription details).",
)

For example, if a user needs medical advice, it transfers them to the Medical Advice Agent.

So, what happens at the code level? — Here’s how the medical agent is defined:

medical_advice_agent = Agent(
    name="Medical Advice Agent",
    instructions="You are a medical professional. Provide general medical advice based on symptoms described.",
)

And this is why the triage agent might transfer a user query to the medical agent.

# Handoff function
def transfer_to_medical_advice():
    return medical_advice_agent

# Adding handoff function to the triage agent
triage_agent.functions.extend([
    transfer_to_medical_advice  # other transfer functions
])

Similarly, if a user wants to book an appointment, the Triage Agent transfers them to the Appointment Scheduling Agent.

Here’s the code behind this transfer.

appointment_scheduling_agent = Agent(
    name="Appointment Scheduling Agent",
    instructions="You are an appointment scheduler. Help users schedule healthcare appointments. You can book appointments using the book_appointment function. If the user asks for 'next Wednesday', you can use that directly as the date parameter.",
    functions=[agent_book_appointment] 
)

Note — The Appointment Scheduling Agent comes with a custom function called agent_book_appointment, which allows it to book slots seamlessly.

The entire app was built in Databutton, using prompts to generate all the backend and frontend code using the Databutton AI App Builder.

Check out this video tutorial below to see how it’s done! 🙌🏽

Update : Just made a live stream on how to build AI Agent System Apps with Phidata and Databutton. You can watch it here: https://youtube.com/live/GiAqN402Kgs

Subscribe to my free weekly newsletter : Ship with AI

What I Built: A Multi-Agent AI Chatbot App using Databutton and Swarm was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

Thinking about your next SaaS idea? Here’s how I went from idea to MVP in just a few hours.

Here’s an overview of what I’ll cover in this post:

• Why this idea clicked for me
• How I started building it
• Where I’m at right now and whether I’ll actually launch

Getting SaaS ideas + building isn’t easy. Period.

I’ve been down this road before — like with my last Micro SaaS app, Fastdeck, where I quickly built an MVP that ended up with over 200+ users trying it as of now.

How I Created a Micro SaaS from Scratch in Just 7 Hours

I often experiment with a quick MVP that might work (well, not always); it’s all about trying and validating as an initial bet.

Idea and Background:

OK, app idea (hmm…) — coming up with a SaaS idea is the toughest part, in my opinion.

Reddit has always been a great source.

And, building and validating what might work has always been the best proven way forward.

So this MVP wasn’t entirely my idea… I found my colleague discussing it over our internal Slack channel, and later he posted about it.

let me share the link of this post below,

Viral Shah on LinkedIn: My friend Mickey Mangan is one of the most inspirational people I know...

“Could a feature be added where the webcam on the device actually records a new video to be uploaded?”

Pretty interesting, isn’t it?

Well, I haven’t exactly gone with that same topic though — instead, I had something similar in mind using video → audio transcriptions as the core feature and a bit ahead from there to “package as a sellable product 💰”

Before I share the app MVP I built, let me write down why I went for this one among so many other ideas:

• Well, it can be a cool product that solves real problems (and down the line, many struggles)
• Plus, OpenAI coincidentally launched their real-time API just in time (read more here).

The second point definitely opens up massive scope. If this idea doesn’t work, I can always pivot with similar features in.

OK, so this is the SaaS idea —

Recruit Swift → Send AI Agents for your next recruit 🫡

Now, how I started building an MVP from this idea

Started by chalking down my idea on a canvas. As always, I went for tldraw.

While planning the first MVP, I wanted to make sure that I understood the Ideal Customer Profile (ICP) and potential value it would provide.

Next, I’m thinking of the core features for my MVP which I can build quickly (and also test the possible hurdles while building):

1. Dashboard for Creating Interviewing Agents and Report Management: let’s say, we add a centralized dashboard available for recruiters/HR to generate a) AI-assisted interview questions and b) deploy ‘recruiting agents — aka Agent Rs’. Of course, the entire process has to be powered by LLMs
2. Video/Audio to LLM-Generated Report: The deployed ‘Agent R’ is sent to potential candidates (with an unique link) to conduct the interview, converting video to audio, then transcribing audio, and finally generating a report using an LLM. The generated report would be again visible in the dashboard for recruiters to evaluate further

I have already shown a live, scratchy demonstration of building this second part over a quick YouTube recording.

The functionalities remain same, however there are still a lot of things that needs to be refined and improved.

Let me add below a quick sketch of Interview board’s app architecture from where it all started.

• We need to build a simple UI with sets of questions that would be generated.
• Once the interview is recorded, an API would extract the audio, generate a transcript using an audio-to-text model, and finally use a text-to-text model to generate complex reports.

Note this part isn’t that straightforward when we start making it interactive for each individual question. Also, we need to hook in a database to store all this data and ensure proper synchronization. This has been one of the challenges as I started to work further on this MVP.

However, for a proof of concept, it worked well.

But the key part, is the huge potential to introduce the real-time API. In my opinion, this would bring a massive impact to the product.

So, where am I right now with this MVP?

☑️ Dashborad for building the Recruiting Agent and getting a deployable link.

☑️ Making the Deployed Agent conduct interviews.

How I built this

The MVP most importantly required a database to work with.

I’ve used Firebase for this purpose.

The features — both the backends and UI, were all built using AI. I used Databutton, an AI app builder, to generate this MVP quickly! (Disclaimer: I’m the community lead for Databutton).

It took me around 16-18 hours to get here ( including the video demo time, after which I revamped the entire app).

What’s left before launch is to connect all these pieces together — i.e. the Interview board and Dashboard. Set up a landing page up and pricing.

Once that’s done, I’ll throw it out to some test users and then go for a small launch, maybe via a social media post or a YouTube demo. You might even see a second blog post here on Medium 🖤

Subscribe to my free weekly newsletter : Ship with AI

I was thinking about an AI SaaS idea and came up with this. was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

Can’t read this? Use our friend link.

Kristoffer Steen was living in a tiny apartment with no space for a wardrobe when he decided to go and solve his own problem by building something from scratch. From then on he gained a passion for helping people in his situation. He started his business, Cutfab, with one CNC machine, a computer with the free 3D design software, SketchUp, and no prior knowledge of furniture design. And he’s been making customers happy for nearly a decade.

But there’s a catch: the Kristoffers of the world simply couldn’t afford his services.

“Our customers today have a lot of space. They can basically buy whatever, and anything fits in their homes. But the people that live in small spaces, they need our services way more, but they can’t afford them.

- Kristoffer Steen, Cutfab & Volva.ai

However, with the right digital customer experience, he’s making tailor-made wardrobes accessible to the masses at a reasonable price.

Half the cost of tailor-made wardrobes is related to measuring, designing, and deciding.

Kristoffer has brought Volva to the world — an AI-based wardrobe configurator to drastically reduce the design cost associated with custom furniture. But how did he get there?

From professional golfer to tailor-made furniture

Kristoffer definitely took a non-linear path to arriving at his current company, inspired by solving problems and building a business the whole way. He started out as a professional golfer in his early years but discovered his passion for creating and building furniture when running into his own problem. After learning about CNC machines, he saw how a professional, data-driven shop mixed with classic craftsmanship could deliver a totally new offering into the market.

Most importantly, he acted on his belief that tailoring your space to your needs and personality improves your productivity, motivation and happiness.

Starting off alone, he’s operating with over a dozen employees with expertise in building high quality wardrobes, cabinetry, and more. His passion for quality and the process shines through when visiting his shop.

The new opportunity with AI

Kristoffer had long wanted to bring his services to people like him when he had his first need. The high time and labor cost related to measuring, designing, and deciding on furniture accounts for nearly half the price of custom.

He’s taken several shots at building a “wardrobe configurator” — an easy-to-use tool that allows individuals to design and request what they want, but it’s never quite delivered. Either it was too hard to use or didn’t generate the right files to actually build things.

He came to Databutton with the thought that AI, particularly large language models (LLMs), could bridge this usability gap. We instantly realized we were kindred spirits, us using LLMs to lower the cost and close the usability gaps for custom apps, and Kristoffer for custom furniture.

We were instantly excited to help him make his app reality and bring this new experience to his customers.

Fun Fact: Wardrobes are generated using a single prompt sent to GPT-4o.

Throughout the process, Kristoffer has been met with tremendous positive feedback from users and customers. So what are you waiting for?

Give Volva a try

If you go through the flow at Volva.ai, you can see how easy it is to design and tweak wardrobes with natural language already. What’s incredible is that these are real SketchUp renderings, with the information needed to actually build and deliver them. Volva is only available in the Norwegian market for now.

Achieve your own mission

We’re incredibly proud to be Kristoffer’s service of choice for building his app and achieving Volva’s mission. Need help with your own app or software? You can also sign up for Databutton or reach out to us with your idea at hi@databutton.io 😊

Kristoffer Steen is going after IKEA’s 47 billion euro revenue with a totally new AI offering was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

A few weeks ago, I had a chat with Sacha Dumay, co-founder of ChatNode, a micro SaaS that went from idea to six-figure exit in under twelve months!

“… it is impossible to learn how to sell a business until you actually do it …you cannot find this information anywhere…”

Sacha shared his learnings on what it takes to create, launch, grow and exit with a successful SaaS product.

1. Get your product out there fast and Iterate

The ChatNode team built their MVP in just two months and launched it on Product Hunt. ( link to their PH launch )

“Speed is important, so let’s do it quick…Don’t waste time on finding the perfect Idea…just build it.”

“You should do whatever work-around or shortcut you can take to ship it …”

Focus on the essential features, launch quickly, and start gathering user feedback.

The sooner your product is live, the sooner you can learn and improve.

2. Don’t Over-Engineer while building your MVP

ChatNode’s initial version was rough around the edges but functional.

“Avoid over-engineering…it’s wasting your time…if there is a bug…they will give us feedback… we fix it”

“… dirty code … it was not like 100% reliable … it works! it works and it was good enough …”

Don’t waste time trying to make your code perfect.

Instead, focus on launching a functional product.

Users will provide feedback that eventually helps in improving the product, so prioritise in getting your MVP in front of them.

3. Pricing Matters

ChatNode launched with a pricing plan, even if it was basic.

“…we’re here to make money, right? You need to have a pricing…”

“Never launch…just a free version. It’s a losing game.”

Always have a pricing strategy from the start.

Even a basic pricing plan is better than offering only a free version.

Free users won’t pay the bills.

Pricing should reflect the value your product provides.

4. Leverage Launching platforms

ChatNode’s initial launch on Product Hunt brought significant traffic and early users.

“Choose your launch day wisely …Weekdays, especially Tuesday, tend to have higher traffic. Plan your launch accordingly.”

Next was AppSumo, which brought in massive inbounds.

“To give you an idea, we made more than six figures net profit in less than two months in AppSumo.”

Plan your launch carefully!Use platforms like Product Hunt and AppSumo to gain exposure and attract users.

Choose strategic days to maximize traffic and engagement.

5. Summer Slump is Real

After the initial buzz, ChatNode experienced a plateau during summer.

“ … In summer you’re going to waste money and time…prepare like plan for September.”

“Summer is really bad for B2B…if summer is flat…just keep building, take some vacation, and plan for September.”

It’s OK to take it slow during the summer!

This time is for planing and preparing for a strong push in September.

Understand business cycles and make strategic decisions about marketing and development.

6. Listen, Learn, and Iterate.

Learning from early users is crucial, but getting feedback can be challenging.

How to solve this?

“…the best way…live chat on your website…it’s very hard…to get feedback…people don’t give…time on it.”

“You get real users…they’re going to motivate you…going to say oh I want that…this is bug here.”

Users appreciate being heard.

Their insights are a goldmine for improvements.

Implement live chat on your landing page to make it easy for users to share feedbacks.

Listen, learn, and iterate.

7. Prepare for Scaling

For ChatNode, the AppSumo launch brought an unexpected surge in users.

“Our infrastructure was pushed…we went from … a thousand users…but we adapted quickly.”

Be ready to scale your infrastructure when opportunities arise.

User spikes can happen unexpectedly, so ensure your system can handle rapid growth.

Adapting quickly to meet demand and maintain the user base is key.

8. The Value of Outsourcing

Sacha learned the hard way that managing customer support and development simultaneously is tough.

Outsourcing became one of his top learnings.

“… as a Founder I never thought customer support would take my time … it’s better to hire someone and train this person and so then … she can or he can take care of that and you can focus on the product …”

Delegate tasks like customer support early on.

This free up the time for product development and strategic growth.

Outsourcing allows you to focus on what matters most!

9. Understanding the AI Market and Choosing the Right Tech Stacks

Selecting the right technology and understanding the latest AI stacks was crucial for ChatNode.

ChatNode used Retrieval-Augmented Generation (RAG). To build RAG, they experimented with different large language model (LLM) providers to find the right fit.

“Everyone worries about the cost of AI…it’s expensive…no, it’s not true…it decreased by 20 or 50 times…don’t worry too much about the cost…”

Stay informed about AI trends and choose technology that aligns with your product goals.

Don’t let cost deter you; AI costs have decreased significantly, making advanced tech more accessible.

More about SaaS building and my personal experience below —

• How I Created a Micro SaaS from Scratch in Just 7 Hours
• Building an AI Micro SaaS Product in Public— Learnings and Updates
• Idea → Build → Deploy. Can We Build SaaS Faster with AI ?

Watch the full conversation with Sacha here —

Let me know what you all think.

I would love to share more stories on SaaS building and the journey from idea to success.

If you’re interested in hearing about more experiences like Sacha’s, leave a comment or get in touch!

Subscribe to my free weekly newsletter : Ship with AI

9 Tips I Learned from an AI Micro SaaS Founder Who Sold it for Six Figures was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

Here’s a friend link for you to read this blog in full, even without a Medium subscription.

How It All Started

I have been using PowerPoint and Google Docs regularly since my academic days.

For the past 2–3 years now, I’ve been using Canva quite a lot.

With the integration of AI, there’s a tremendous opportunity to create slides in a short amount of time. Yes, most of the aforementioned tools have already integrated this feature to some extent.

But what if we make it even simpler and goal-oriented ?
What if we could create slides via a prompt with minimal features? Keeping all the flashy features out of sight and concentrating on the core function.

Idea : Prompt → Slide deck Generator

That’s it. No need to worry about design or background — all done by AI.Just prompt and get your decks ready in a minute or two. Next, you have the slide decks ready to share.

That’s how FastDeck AI started.

However, it’s now more important to validate the idea and just ship it to a couple of early users.

It worth to mention that I was very much inspired by Marc Lou and his approach to SaaS development.

“ … Instead, think in weeks, and ditch ideas that take longer than a month. You can also lower your time-to-market:
- Presell a product
- Ditch code testing & dark mode
- Break big ideas into 1-feature-only products
There is no perfect idea.

The only way to know is to launch it.

This framework helps you increase the odds of success and get profitable faster… ”

How to validate a startup idea

Building FastDeck AI

I needed to use a tech stack that I’m well-versed in to build and prototype this product quickly. I chose Databutton.

Disclaimer: I work for Databutton and have been using this tool extensively over the last couple of months.

The Databutton AI app developer specialises in building full-stack web apps and MVPs conversationally, especially when the apps are AI-driven themselves.

With Python as the backend, React as the frontend, and one-button deployment, Databutton serves as a good starting point to build your application.

A Quick Sketch of the App Architecture

Spent the first couple of hours on sketching the main app architecture and setting up the key components:

• How the frontend would look
• Backend/API calls to make from the UI
• How the app would interact with the end-users

Also, I made sure to document each of the processes and share while I build.

Here’s the first tweet where I could make an AI Agent create a slide from scratch…

Iterating and Building the Minimal Features

Okay, we have something up!

The next step is to iterate further and get the minimal features app ready as early as possible.

I prioritzed next on the app’s UI more.

• Added some headers
• Polished the UI
• Made the “Generate Deck” button visible
• Changed the Background theme
• Added a “Logout” Button

Getting the Backend Work Done

With the app’s UI taking shape, it was time to focus on the Python backend and improve the main backend code to ensure smooth functionality and performance.

Implementing Core Functionalities:

• Firestore for User Authentication: Leveraged Firestore for managing user authentication. Databutton’s template was particularly helpful in setting this up efficiently.
• Agent Code Generation: Used a simple prompting hack with OpenAI’s GPT-4 to generate code faster. This approach worked remarkably well for speeding up development.
• Authentication + Firestore Database Setup: Spent time deciding on the best way to set up authentication and structure the Firestore database. Also, finding a scalable architecture that could grow with the app’s user base.

Key APIs (Backend) Necessary for the quick MVP

agent API: The core backend API responsible for creating slide decks based on the user prompt. Generates the slide decks and passes them to the frontend for display.

firestore_database API: Manages all user authentication and stores generated decks on a per-user basis.

Work in Progress: stripe_webhook API: Currently integrating payment functionality using Stripe. This will handle payment processing and subscription management

Polishing the MVP & Testing it Before Sharing with the World

This time, I focused on adding more low-hanging functionalities.

Features like the number of slides that can be generated or the tone of the slides (e.g., professional, casual) were introduced. These parameters are passed to the backend as part of the prompt.

Next, tested the MVP thoroughly to ensure all features worked as expected with the least number of bugs possible. I verified that the slide generation, user authentication, and other core functionalities operated smoothly.

Oh wait…

Also realized the need for a landing page to demonstrate all the possible features of FastDeck AI.

So, quickly generated an informative landing page that highlights the app’s key functionalities. Additionally, recorded a short video explaining how the app works.

More Promotion Over Social Media (X, LinkedIn, and TikTok) and Gathering Early Users’ Feedback

Only failed once, so I reduced the slides to 1; this is so much fun!!! and AMAZING!! nice work!

User feedback helped quite a bit! Based on the feedback, I added a new feature to share decks directly from the app.

In addition to social media promotion, I sent a few cold emails to early users.

How It Looks Now and What’s Next?

Now, we have more than 30 active users using FastDeck AI.

Somewhere along the way, I lost the momentum to promote it lately, but seeing the constant signups and positive feedback from the early users has been incredibly encouraging.

Yes, I would definitely love to work on FastDeck AI further. So, would love to hear your feedback and improve it!

Try it out and let me know what you think.

FastDeck AI

And most importantly, thank you for following along on this journey.

Stay tuned for more updates and improvements!

Similar posts you might like to read —

• I was thinking about an AI SaaS idea and came up with this.
• Building an AI Micro SaaS Product in Public— Learnings and Updates
• 9 Tips I Learned from an AI Micro SaaS Founder Who Sold it for Six Figures

How I Created a Micro SaaS from Scratch in Just 7 Hours was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

The market for niche AI-powered products is growing considerably.

Software products can now target narrower audiences with more specific problems in a profitable way, just like TV and entertainment went from mass-appeal to narrower content targeting smaller audiences in the past 15 years. Why? The reason is slightly different for the two different industries, yet still relevant.

New companies can empower people with specific AI experiences

Entertainment as an example for software

In the entertainment world, new means of distribution via streaming made it possible to create content that hits a much narrower segment than in the era of classic TV programming, while still earning money. What did streaming bring?

• A new way to consume with on-demand
• A new form for a catalog with infinite shelf-life and access
• Much better audience data and ability to target

Netflix series like Master of None and Never Have I Ever never would have seen the light of day on broadcast TV. These are stories about experiences and life as an Indian American, while eventually gaining further appeal, had a narrower initial audience.

For SaaS and software products, new means of production via AI code generation brings the possibility to target narrower segments and problems. What does AI code generation bring?

• Lowers the cost to create something custom, tailor-made (both in time and money)
• Gives unprecedented control to people with a business idea, compared to traditional outsourcing or freelancers

Awareness and adoption lags behind tech

It’s important to remember that in-market technology does not equate to market awareness! Niche products can outcompete general purpose technology like ChatGPT in two ways:

• A fit-for-purpose product experience and workflow
• A positioning and go-to-market that targets the customer problem

People that are deep into the AI wave will be aware of new things the latest and greatest LLM is capabile of. But again and again we see that a large majority of people don’t even know what is possible (and honestly, most don’t even know what an LLM is).

A great example is Elle, who I spoke to in the video above. She discovered how useful ChatGPT was for her ADHD, and built https://neuromermaid.ai/ to help people like her with researching and learning.

What problem are you looking to solve? Use Databutton’s AI agent to build your next product!

AI code generation creates the opportunity for a new wave of niche software products was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.

Imagine hiring a new sales development representative (SDR) right out of university. Could you provide them with goals and business context, then expect immediate results? While possible, you’re more likely to end up extremely disappointed.

Typically, a significant amount of resources are necessary to get a new employee up to speed. In fact, a Gallup report reveals that it takes around 12 months for a new employee to become fully productive.

The onboarding process often includes:

• A detailed job description outlining clear expectations, responsibilities, and tasks
• Regular check-ins, guidance, and course correction with the manager
• Shadowing sessions to learn from others’ success
• An onboarding buddy for asking any questions
• Explicit training sessions and materials on job execution

This article from HBR offers 14 suggestions for new managers on how to onboard new employees effectively.

While expectations for human labor are clear, the market still doesn’t quite know what to expect from AI agents. AI agents are autonomous systems performing specific tasks without human intervention. And AI agents are taking the startup world by storm. Y Combinator has 67 startups that are listed in this space as of May 2024. The problem is, building just any AI agent and expecting it to succeed can be challenging.

But don’t worry! With the right tools in your toolbox, you can identify a promising market and develop an AI agent that solves real problems. The following discusses a framework for deciding what to build, including examples of successful in-market agents and ideas yet to take off.

A framework for identifying AI startup ideas

To fully grasp the potential of AI agents, we begin by examining the key organizational needs within a company. The simplest way to identify a valuable AI agent is to consider which roles, responsibilities, and tasks you would delegate to a person. In terms of markets, think about the services or consulting models that are frequently paid for, rather than hiring in-house. This aligns with current needs and purchasing behaviors.

The one-person unicorn

While the idea of AI replacing human labor may seem unsettling, we’ve observed that all technological advances, like the second industrial revolution, have significantly reshaped the labor market. Although many jobs were lost, entirely new ones were created, leading to an economic boom (link).

In some cases, the advancement of mechanization substituted for low-skilled workers altogether. Both the number of unskilled and skilled workers increased, as their wage rates grew. Engineering colleges were established to feed the enormous demand for expertise. Together with rapid growth of small business, a new middle class was rapidly growing, especially in northern cities.

Numerous experts predict the rise of the one-person unicorn, which is the equivalent labor shift going forward. AI agents are what makes this possible.

“In my little groupchat with my tech CEO friends there’s this betting pool for the first year that there is a one-person billion dollar company,” Altman told Ohanian. “Which would have been unimaginable without AI and now will happen.”

Sam Altman, OpenAI

The framework

Step 1 — Determine your agent’s role
First, consider the expertise you bring to the table. Are you a seasoned marketer, a patent lawyer, or a PR expert? Your goal is to encapsulate this knowledge in an AI agent, tailoring it to your customer’s operational model.

Describe the experience and knowledge that this agent will possess.

Step 2 — Identify your Ideal Customer Profile (ICP)
Which companies and buyers could most benefit from your AI agent? Are they currently underserved? Can you create a new market by offering a unique price point? There is plenty of excellent material available to help define an ICP.

Step 3 — Define your agent’s success
Outline the specific responsibilities and outcomes for which they are accountable. This could include tasks like “successfully submitted patent application” or “contacted 5 publications.”

Step 4 — Determine the tasks and context your agent needs to succeed
Consider what information, context, and tools a human would need to excel at these tasks. Does your agent require a comprehensive handbook? Maybe access to an excellent email client and CRM? Will it need input from its “boss” at times? If so, how often and for what reasons?

Step 5 — Assess feasibility
Is it feasible? If all the necessary data, context, and tools are available via API, the only obstacle is execution (easier said than done, of course). So, go for it!

AI Agent Startups

We now have a way to determine good AI agent startup ideas!

When it comes to agents that are successful in-market, many of them involve roles like software development and design. My core hypothesis as to why this is the case is two-fold:

1. The first people working with new technologies are developers and designers, and therefore the agents being built are based on their expertise.
2. The early adopters for technology like this are tech-forward individuals whose needs often involve building digital products.

In-Market AI Software Developers

Devin made a big splash as the “world’s first software developer” on March 15, 2024. While being a very big and inspiring vision, Devin is not yet available in-market. The scope of what a software developer is responsible for is immense, and it’s quite hard to cover such a spectrum of tasks and requirements with an AI agent, especially today.

The vision is for building a general purpose “AI software developer” is large and therefore it’s naturally hard to project and see how AI could fill those large boots. After all, a software developer doesn’t simply produce code, she needs to prioritize work, fix bugs, build new features, define how things should work, test, understand customers, collaborate with other roles, and a lot more!

In-market agents narrow down the role, tasks, and target market in order to start out with a niche. What is key for these agents is that they codify knowledge about their respective fields (app development, bug fixing, and frontend development) as well as have the right tools to execute on their tasks.

Tusk — AI-created pull requests for annoying tickets

Sweep — AI Junior Developer that maintains your legacy codebase

Magic Patterns — Your new frontend assistant.

…and there are many more out there today for a variety of development needs.

Proposed AI Marketing Analyst Agent

Now let’s use the framework to develop a proposal for a competent AI agent that a skilled digital marketer could deliver based on their expertise. This can be applied to other marketing roles as well. Let’s try to keep it narrow in order to build an initial niche for the agent.

The Agent’s role
You’re a digital marketer with expertise in driving the success of PPC (Pay-Per-Click) campaigns across multiple platforms including Google, YouTube, Instagram, TikTok, Facebook, and LinkedIn. You will be instrumental in managing, building, and optimizing online advertising efforts to achieve the best possible ROI. Your core responsibility is to prepare detailed reports on campaign performance and recommend cross-channel optimizations.

The Ideal Customer Profile (ICP)
SMBs with 200–1000 employees, currently transitioning to e-commerce as a new sales channel.

Success for the agent
Significant increase in impressions and revenue from PPC campaigns.

Determine the tasks and context your agent needs to succeed
The agent would need access to all ad systems and data, access to a business intelligence or reporting tool to generate reports (e.g. Tableau) and context on performance expectations, the company’s products, and more.

Feasibility
Since all necessary resources are available via API and text input from a potential customer, it seems like an excellent opportunity to disrupt the market with an AI agent!

Proposed AI UX Researcher Agent

Let’s apply the same framework used for a digital marketing agent to develop a specialized AI agent that a seasoned UX researcher could create based on their expertise. This agent will focus on enhancing user experience insights, particularly in digital environments.

The Agent’s Role
You’re a UX researcher with deep knowledge in conducting and analyzing user behavior studies and usability testing across digital platforms. Your principal responsibilities include synthesizing complex data into actionable insights, creating user personas, and suggesting UI/UX improvements based on user feedback and behavioral patterns. You will assist teams in understanding what drives user engagement, satisfaction, and retention.

The Ideal Customer Profile (ICP)
Tech startups and mid-sized tech companies that are actively developing or optimizing digital products, such as mobile or web apps, and need to enhance user experience to increase adoption and satisfaction.

Success for the Agent
A measurable improvement in user experience metrics, such as increased user retention or higher engagement rates.

Determine the Tasks and Context Your Agent Needs to Succeed
The agent requires access to user interaction data, feedback channels, analytics tools, and design software. It needs context regarding the target user demographics, business goals, existing UX pain points, and design standards. Integration with platforms like Mixpanel, Heatmap tools, and A/B testing software would be crucial for in-depth analysis and reporting.

Feasibility
Given the digital nature of UX research and the availability of APIs to access necessary user data and analytics tools, there is a strong opportunity to develop an AI UX Researcher Agent. The technology required to analyze and interpret user behavior and feedback digitally is well-established, making this a viable AI agent that could significantly impact product development processes.

Final thoughts

AI agents are not just about technology; they involve codifying real knowledge into working models and identifying the right markets. As we stand on the brink of this game changing shift, startups have the unique opportunity to lead the charge, transforming business challenges into a solution with AI agents.

I’m not in Sam Altman’s betting pool, but seeing the tremendous activity around AI agents (and hopefully even more) makes me believe the one-person unicorn isn’t as far away as we might think.

Thank you for reading. And check out Databutton’s AI app developer for building your big AI startup!

How to Identify Your Breakthrough AI Startup Idea was originally published in Databutton on Medium, where people are continuing the conversation by highlighting and responding to this story.