An AI-powered app with zero lines of backend code

When I first set out to build Orderly I had checked most of the boxes the YCombinators of the world tell you to:

• Founder market fit? Check (I've been working in returns technology for years)
• Market validation? Check (A list of verbal commits and waitlist signups)
• Solving my own problem? Check (Inundated with wedding gifts we hadn't figured out how to return)

Why do those matter? Because I wanted to build an app that made returns manageable for consumers. It would connect to your email to identify all of your online orders, track where they were, use retailers' returns policies to remind you to make returns, and even allow you to schedule home pickups for your returns.

Unfortunately, the box I didn't check was the ability to build it myself...or knowing somebody who could build it for me... or even being able to find a reputable stranger to build it for me (keyword is reputable. lots of strangers wanted to build something for me).

I need to start doing something though, so purely out of desperation I decided to see how much I could actually get done myself.

Step 1: Literally where do I start?
So turns out software needs a frontend and backend. Thought you all should hear it from me first.

Logically, if you're going to do one of those first backend makes the most sense, so I decided we'd start there, which sounds smart and reasonable until you realize you don't know how to stand up a database or build an API or anything.

Step 2: I’ll try anything
I might be the first person in history to shoutout Reddit Ads, but that's what led me to the first piece of tech that made this journey possible: Xano.

Basically, Xano is a no-code backend tool. It allows you to drag and drop GUI blocks that represent code and add configurations to them, but more importantly (at least for me) it does all of the work of standing up a database and APIs for you.

For someone with a data science background that understands how algorithms are constructed and basic data structures this was essentially all I needed. So I went ahead and set up all my tables and endpoints - feeling great. Even connected to the Gmail API to get content from user messages - incredible. I had essentially accomplished what I thought would take 2 weeks in 24 hours.

Felt less great when I realized I now needed to build an algorithm to extract all of this data from user emails.

Step 3: Oh, this is kind of working
I figured there must be some semi-standard way retailers send order confirmation emails, so I started building basic regex based on the order confirmation emails in my own inbox. That went...poorly.

Turns out retailers couldn't care less about standardizing across the industry. In fact, they're trying their hardest to have their emails stand out. As such, I needed something that can take a lot of differently structured text inputs and output them in a consistent (JSON) structure.

Enter ChatGPT - specifically, OpenAI’s chat completions endpoint. I half-jokingly just fed it the text from an email and gave a lazily constructed prompt along the lines of “tell me the retailer name, order number, and price outputted in JSON format.” It was god damn magic. Came out perfectly.

Example using ChatGPT of what my API prompt would look like:

I had pegged 3-4 weeks for devising an NLP model to extract the relevant details from emails. Instead I spent about 48 hours fine tuning my prompts and my backend was in essence MVP-level complete.

Step 4: ChatGPT, the gift that keeps on giving
As I revealed to you all earlier, software often requires a frontend. I did a couple of spikes on building with no-code frontend tools (namely Adalo and Bubble.io). They were way more restrictive than online reviews had led me to believe. Could barely handle my Google OAuth flow that requested one permission more than the typical set.

So again I was left with the choice of hiring a dev or going at it myself (I didn’t know any frontend languages, like at all). I decided I’d get started trying to learn React Native, and that I’d use ChatGPT as my tutor.

What happened next was again, god damn magic. I fed it a sample of my order data and asked it to give me an outline of how I’d build a screen that listed out people’s orders. Instead of it answering my prompt as I expected, it literally just…built the screen? And it worked?

So I did that again. And again, and again, and again for all the different functional components I needed in my frontend. It even taught me how to set up Expo, which has been a saving grace. Expo is essentially a react native SDK that abstracts away the complex programming details of testing and building apps. Highly recommend it.

Step 5: Wow, ok, so we did it
Yes we did. In October 2023 I released a pretty rough and dirty MVP to some close friends and family. In the following weeks I talked to these users daily, iterated, and iterated, and iterated, until we had what I felt comfortable releasing to the world. Which is what I did, in November 2023 the week of Thanksgiving / Black Friday (the beginning of peak season for returns).

Today, Orderly still runs on the same technology (Xano, OpenAI, Expo) I outlined above, even as its scaling up and up and up to hundreds and thousands of users. We're live on both app stores, ingesting, processing, and helping users with over 5,000 orders and returns per month.