Demystifying AI Applications – How to Create AI Applications using RAG and Predictive Models
As AI gains momentum, more business leaders are asking about AI applications. In most industries, the implementation of AI seems inevitable, but that does not mean they should be implemented without scrutiny.
Leaders in every industry have many questions, including:
- Are they Real?
- Can they be built without massive budgets?
- Do they actually work?
At the same time, I sometimes encounter the opposite view — that AI has almost magical powers While AI can feel magical, it is still rooted in very tangible things: data, business rules, probability, and statistics.
The purpose of this article is simple:
- To explain, in plain terms, how an AI application is built — and to show that AI is tangible, accessible, and achievable for most organizations today.
- I’ll use a real example that combines Generative AI and Predictive AI to solve a real business problem we’ve encountered at Apps Associates.
The Business Problem
Recently, the Apps Associates team worked with a construction company that wanted to streamline their processes by predicting future project costs based on historical projects.
All of the relevant data — project characteristics and detailed cost breakdowns — was stored in PDFs. The information was there – years of pricing data at their fingertips – but it wasn’t structured in a way that a machine learning model could use.
To help them leverage the valuable data in PDFs, we decided to build an AI application that combined Retrieval Augmented Generation (RAG) and Predictive Machine Learning.
Let me now define these concepts in simple, non-technical terms.
What Is Retrieval Augmented Generation (RAG)?
RAG allows a large language model (LLM) to become aware of your private data and answer questions based on that data. In this case, we made the LLM aware of the company’s PDFs.
What Is a Predictive Model?
A predictive model makes forecasts based on past data.
- To train a predictive model, you provide it with many examples where both the inputs (project characteristics) and the outputs (actual project costs) are known.
- In simple terms, you show the model the answer over and over under different scenarios and, eventually, it learns patterns and can make predictions on new projects.
Creating an AI Application
Now that we’ve covered the basics of RAG and Predictive Models, I’ll explain how we solved this particular business problem using a custom AI application.
Step 1: Turn PDFs into a Training Dataset
All the information we needed was inside the PDFs:
- Project attributes
- Cost categories
- Total project costs
The first goal was to extract that information and organize it into a training dataset (a structured file the model can learn from).
Here’s what we did — intentionally described in non-technical terms:
- Collected and organized the PDFs to ensure we had enough historical examples.
- Extracted relevant text and numbers from the PDFs.
- Stored that information in a searchable format.
- Used RAG techniques to assemble the extracted data into a structured training dataset.
At this point, we had converted unstructured PDFs into structured data suitable for machine learning. If you’re curious about what makes data suitable for an AI model like this, check out another one of my blog posts: What does Good Data really mean for AI?
Step 2: Train the Predictive Model
By taking the time to ensure your data set is in a suitable format, you set yourself up for success in using that data elsewhere. Once the data is organized, we:
- Used the training dataset to train a predictive model.
- Evaluated its accuracy.
- Tested different model approaches.
- Refined the dataset where necessary.
Yes, some coding was involved — but the tools today make this far more accessible than many people assume.
The important point: this process is repeatable and something that most IT organizations could handle.
The end result is a repeatable way to extract structured data from PDFS, and a repeatable process to train a predictive model using that data.
Step 3: Put It into the Hands of Users
In this example, our client wanted business users to have autonomy over this model. Users needed to enter characteristics of a new construction project and instantly receive a cost prediction.
To make this the most user-friendly, we built a simple front-end application that:
- Accepts project inputs.
- Calls the trained model.
- Applies it to the new data.
- Returns a predicted cost to the user.
End-to-end, we built this application in a matter of weeks.
The Bigger Message
AI applications are not mystical. They are not reserved for massive organizations with teams of PhD data scientists.
Modern platforms have evolved to the point where most IT teams, technically savvy business teams, and organizations of all sizes can build meaningful AI applications within reasonable budgets and timeframes.
Yes, you need the right skills. But solutions like the one described above are attainable and achievable with AI technology available today.
For Those Interested in the Technology Stack
For this solution, we used:
- Oracle AI Data Platform as the development environment
- Python notebooks
- Oracle Cohere as the Large Language Model (LLM)
- XGBoost for predictive modeling
- Oracle Cloud Infrastructure (OCI) Object Storage to store the PDF’s and the trained model
- Oracle APEX to build the front-end application
Below is a short video demonstrating the end-to-end AI application in action — from user input to real-time prediction:
AI doesn’t have to be theoretical or overly complex. At Apps Associates, we work with clients to thoughtfully apply AI to real operational challenges and deliver tangible results. If you’re considering how AI could support your business objectives, contact Apps Associates today.
