Build a consistent labeling workflow
Llama 3 zero-shot classification maps approximately 75,000 sandbox transactions to 18 spending classes. A TensorFlow classifier is evaluated against logistic regression on a chronological holdout.
All selected workMake spending
Decisions behind
01 02 03 04 0.94 <6%
Independent project / Engineering case study
Meridian.
From transactions to a clearer picture.
An AI personal finance assistant connecting transaction classification, spending forecasts, and retrieval-grounded budgeting recommendations.
01 / Problem
Make spending
understandable.
Individual transactions reveal little on their own. Meridian turns spending history into categories, monthly forecasts, and budgeting recommendations grounded in each user’s transactions. Development and classification evaluation use sandbox data.
Implemented scope
Plaid Sandbox ingestion, transaction classification, monthly forecasts, and retrieval-grounded budgeting recommendations, with Python services on Lambda and ECS. Docker Compose supports local development, alongside the AWS delivery workflow.
My contribution
I built the classification and retrieval workflows and connected them to Python services on Lambda and ECS. Airflow orchestrates Plaid ingestion, MLflow tracks experiments, and Docker, Terraform, and GitHub Actions support deployment.
02 / Engineering decisions
Decisions behind
the system.
Ground recommendations in personal history
The RAG workflow retrieves context from each user’s transaction history to generate budgeting recommendations tied to their spending.
Forecast recurring spending
Prophet forecasts monthly spending in stable categories such as utilities and groceries. High-variance discretionary spending is excluded from the forecast.
Connect experiments to deployment
MLflow tracks model experiments. Docker Compose supports local development, while Terraform and GitHub Actions support delivery of Python services on AWS.
03 / Results
Results
Classification on sandbox transactions and monthly forecasts for stable spending categories.
Macro-F1 · classification
TensorFlow versus a 0.82 logistic-regression baseline, evaluated on a chronological holdout from approximately 75,000 sandbox transactions. Llama 3 zero-shot labels cover 18 spending classes.
MAPE · monthly spending forecasts
Prophet forecasts recurring monthly spending in stable categories, including utilities and groceries. This result excludes high-variance discretionary spending.