Building a Financial Assistant

Think Different - Dhiraj Patra
4 min readNov 25, 2023

Power of 3-Pipeline Design in ML: Building a Financial Assistant

In the realm of Machine Learning (ML), the 3-Pipeline Design has emerged as a game-changer, revolutionizing the approach to building robust ML systems. This design philosophy, also known as the Feature/Training/Inference (FTI) architecture, offers a structured way to dissect and optimize your ML pipeline. In this article, we’ll delve into how this approach can be employed to craft a formidable financial assistant using Large Language Models (LLMs) and explore each pipeline’s significance.

What is 3-Pipeline Design?

3-Pipeline Design is a new approach to machine learning that can be used to build high-performance financial assistants. This design is based on the idea of using three separate pipelines to process and analyze financial data. These pipelines are:

The data pipeline: This pipeline is responsible for collecting, cleaning, and preparing financial data for analysis.

The feature engineering pipeline: This pipeline is responsible for extracting features from the financial data. These features can then be used to train machine learning models.

The machine learning pipeline: This pipeline is responsible for training and deploying machine learning models. These models can then be used to make predictions about financial data.

Benefits of 3-Pipeline Design

There are several benefits to using 3-Pipeline Design to build financial assistants. Some of these benefits include:

Improved performance: 3-Pipeline Design can help to improve the performance of financial assistants by allowing each pipeline to be optimized for a specific task.

Increased flexibility: 3-Pipeline Design makes it easier to experiment with different machine learning models and algorithms. This can help to improve the accuracy of financial predictions.

Reduced risk: 3-Pipeline Design can help to reduce the risk of financial assistants making inaccurate predictions. This is because the different pipelines can be used to check each other’s work.

How to Build a Financial Assistant with 3-Pipeline Design

The following steps can be used to build a financial assistant with 3-Pipeline Design:

Collect financial data: The first step is to collect financial data from a variety of sources. This data can include historical financial data, real-time financial data, and customer data.

Clean and prepare financial data: The financial data must then be cleaned and prepared for analysis. This may involve removing errors, filling in missing values, and normalizing the data.

Extract features from financial data: The next step is to extract features from the financial data. These features can be used to train machine learning models.

Train machine learning models: The extracted features can then be used to train machine learning models. These models can then be used to make predictions about financial data.

Deploy machine learning models: The final step is to deploy the machine learning models into production. This involves making the models available to users and monitoring their performance.

Understanding the 3-Pipeline Design

The 3-Pipeline Design acts as a mental map, aiding developers in breaking down their monolithic ML pipeline into three distinct components:

1. Feature Pipeline

2. Training Pipeline

3. Inference Pipeline

Building a Financial Assistant: A Practical Example

1. Feature Pipeline

The Feature Pipeline serves as a streaming mechanism for extracting real-time financial news from Alpaca. Its functions include:

- Cleaning and chunking news documents.

- Embedding chunks using an encoder-only LM.

- Loading embeddings and metadata into a vector database (feature store).

- Deploying the vector database to AWS.

The vector database, acting as the feature store, stays synchronized with the latest news, providing real-time context to the Language Model (LM) through Retrieval-Augmented Generation (RAG).

2. Training Pipeline

The Training Pipeline unfolds in two key steps:

a. Q&A Dataset Semiautomated Generation Step

This step involves utilizing the vector database and a set of predefined questions. The process includes:

- Employing RAG to inject context along with predefined questions.

- Utilizing a potent model, like GPT-4, to generate answers.

- Saving the generated dataset under a new version.

b. Fine-Tuning Step

- Downloading a pre-trained LLM from Huggingface.

- Loading the LLM using QLoRA.

- Preprocessing the Q&A dataset into a format expected by the LLM.

- Fine-tuning the LLM.

- Pushing the best QLoRA weights to a model registry.

- Deploying it as a continuous training pipeline using serverless solutions.

3. Inference Pipeline

The Inference Pipeline represents the actively used financial assistant, incorporating:

- Downloading the pre-trained LLM.

- Loading the LLM using the pre-trained QLoRA weights.

- Connecting the LLM and vector database.

- Utilizing RAG to add relevant financial news.

- Deploying it through a serverless solution under a RESTful API.

Key Advantages of FTI Architecture

1. Transparent Interface: FTI defines a transparent interface between the three modules, facilitating seamless communication.

2. Technological Flexibility: Each component can leverage different technologies for implementation and deployment.

3. Loose Coupling: The three pipelines are loosely coupled through the feature store and model registry.

4. Independent Scaling: Every component can be scaled independently, ensuring optimal resource utilization.

In conclusion, the 3-Pipeline Design offers a structured, modular approach to ML development, providing flexibility, transparency, and scalability. Through the lens of building a financial assistant, we’ve witnessed how this architecture can be harnessed to unlock the full potential of Large Language Models in real-world applications.



Think Different - Dhiraj Patra

I am a Software architect for AI, ML, IoT microservices cloud applications. Love to learn and share.