Not all mistakes are bad. Sometimes these can help you make some decisions you’d never think about. 

In this article, I will share my story of starting and selling Docstract, a tool for data extraction from unstructured documents such as receipts, invoices, and so on, through LLM technology, in 3 months. 

Docstract’s Birth

In 2022, I invested in Jobful and started a partnership through my web development agency. Through this, in 2023, we received a request for a plugin solution in Jobful from a Romanian Enterprise Client — a Python microservice, which, by connecting it to Jobful, extracts data from onboarding documents (Identity Card, Birth Certificate, Education Diploma, etc.)

After successfully delivering this project with a colleague from the university, Bianca Necula, I realized that based on this work, we could build a product for unstructured documents targeting small and medium businesses.

We worked on an MVP that had good accuracy. We used a mix of tools like: 

• Laravel & Inertia JS for the Web App;
• Python & OpenAI for the AI part micro-service;
• Google Cloud for the Python microservice;
• AWS S3 bucket for the file storage;
• OVH for the VPS hosting;
• Gitbook for the API documentation;
• Tawk.to for the website chat;
• Mailchimp for the newsletter;
• Cloudflare;
• Stripe for the payments;
• Github for the repo;
• Hubspot for the sales CRM part;
• Apollo for the outreach system;

The mistake

After trying to onboard some clients, we realized that we cut some corners:  we were building a product without validating it in the market.

Only once we had the MVP ready, we started discussing with different clients from the mid-market and realized that the pain was not too hard for this market because they wouldn’t process a lot of documents. We tried to focus on different niches, but the result was the same — the pain point was felt only by bigger companies, and bigger companies need a lot of input and output integrations.

Sources: The Recursive