The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
📄 Building an AI Invoice Details Extractor — Step 1: Data Collection & Preprocessing
As part of my work on an AI-powered Invoice Details Extractor, I’ve completed an important intermediate stage of the project: dataset collection and preprocessing.
For this project, I collected the SROIE (Scanned Receipts OCR and Information Extraction) dataset from Hugging Face. The dataset provides receipt images along with OCR information, word-level bounding boxes, and entity annotations, making it suitable for training document-understanding models.
🔹 Data Collection
Collected invoice/receipt images and their corresponding annotations.
Extracted OCR words and their spatial coordinates.
Identified key entities such as Vendor, Date, and Total Amount.
Organized the data into a format suitable for model training.
🔹 Preprocessing
Processed and normalized the OCR text.
Normalized bounding-box coordinates to a 0–1000 scale.
Converted entity annotations into BIO token-level labels.
Mapped labels to numerical IDs.
Prepared the images, text, bounding boxes, and labels for a LayoutLM-based Transformer model.
Applied tokenization, padding, and truncation to create model-ready inputs.
The goal of this stage was to transform raw receipt data into a structured, consistent, and model-ready dataset, while preserving the relationship between the text and its position within the document.
This preprocessing pipeline forms the foundation for the next stage: training the document-understanding model to automatically identify and extract important invoice information.
🚀 Next Step: Model training and evaluation.
#AI #MachineLearning #NLP #OCR #DocumentAI #LayoutLM #DeepLearning #PyTorch #ArtificialIntelligence #InvoiceAutomation #DataPreprocessing #HuggingFace
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
I built an AI agent that types customer orders into QuickBooks, so you don't have to type.
Many distributors still pay someone to retype every customer order: a PDF purchase order here, an email there, a spreadsheet, even a text message. A part-time order clerk costs about $1,500 a month.
So I built an Order Desk Agent. When an order arrives, it:
reads it in any format (PDF, email, spreadsheet, text)
finds the customer and matches every line to the product list
converts units (each, dozen, case) and uses that customer's contract price
drafts the QuickBooks sales order and the confirmation reply
The important part: nothing goes into QuickBooks without a person's OK. Anything unusual is flagged, not guessed: a price below contract, a quantity far above normal, an item it can't match.
In the picture: a hotel's PDF purchase order (middle) becomes a draft sales order (right). 6 of 6 lines matched, units converted to cases, and 1 price difference flagged for review. Sample data, from my 2-minute demo.
I set it up on your own product list and past orders in about a week. First month free, then $500 a month.
If your team types orders by hand, message me.
Watch the 2-minute demo: https://gntech.gm-tools.workers.dev/order-desk
How does it do with handwritten POs or a photo of one? That's usually where these break.
Designed a modern, responsive AI-powered website for AIVORA, an all-in-one AI productivity platform. The project features an intuitive UI/UX, AI chatbots, AI content generators, intelligent virtual assistants, workflow automation, document analysis, knowledge management, and analytics dashboards. Focused on clean design, seamless user experience, mobile responsiveness, and making powerful AI tools accessible to individuals and businesses.
AI really changed how I work. I spent the last month learning how to use it to amplify my work, now I’m looking for projects where I can put that into practice at scale. Open for work!
A month learning to use AI to amplify your work before taking it to clients is the right sequence, Jacob. Putting it into practice at scale is where the real lessons show up. All the best finding the right projects.