Alan Sammarone

We developed an end-to-end system for bank transaction classification using weak supervision and deep learning.

I was involved in the ideation and research phases and led the productionization effort. My work included designing and implementing the real-time inference architecture, integrating the system with Kafka and KSQL for event batching, and deploying scalable models via Kubernetes and Kubeflow Pipelines.

The system uses heuristics and unsupervised embeddings to generate weak labels, which are then used to train GRU-based discriminative models. Our pipeline outperformed Plaid API in several tasks, achieving over 90% accuracy across nine financial categories. The architecture was optimized for rapidly onboarding new classification tasks with minimal manual effort.

The Python backend for an MVP that leverages three OpenAI capabilities in sequence - gpt-4o-transcribe for converting speech to text, gpt-4o for processing and generating responses (with tool-calling support), and gpt-4o-mini-tts for converting the response back to speech - to provide an interactive voice interface that handles end-to-end customer support tasks without human intervention.