Developed and launched a full-stack LLM Agent app for financial document analysis.

Reliably extracts accurate, complex financial tables from a user’s PDF documents and photos.

Users interact with the Agent to compose the extracted data into financial answers and insights.

Combines the power and accuracy of heavier, more technical tools with the usability and ease of taking a photo or simply uploading a document.

Features and UX developed in tandem with a small business owner and financial analyst.

LLM Agent Backend

Built a custom LLM Agent framework using the claudette library from Answer.AI to fully leverage Anthropic’s prefill, tool-calling, and tool-loop functionality.

Fine-tuned a modernBERT embeddings model with QLoRA, using a contrastive loss, on a curated and cleaned set of financial data.

Developed a hybrid RAG engine by combining fine-tuned embeddings with BM25, with a final re-ranker stage to maximize the quality of our Agent’s context.

System continuously tested against an internal set of complex documents (financial reports, medical tests), routinely showing better performance than off-the-shelf LLMs.

Pioneering LLM Applications

Built a complete RAG application for the first internal LLM hackathon in the Fraud organization.

Led a hackathon team to 2nd place building a Graph-RAG agent based on Knowledge Graphs for internal documentation.

Consulted on general Deep Learning and ML expertise for team members across the organization.

Fraud Detection Model Deployments

Created the guidelines for migrating python Fraud deployments to an internal Kubernetes fork.

Migrated five production fraud detection systems from Databricks to Kubernetes.

Established and shared best-practices around Jupyter notebook development for PySpark models.

Improved the accuracy, reliability, and speed of pipelines dealing for large-scale credit card features.

Daily monitoring and triage of three critical production Bayesian Graph fraud detection systems.

Semantic Forensics (DARPA SemaFor)

Developed and trained convolutional neural networks to classify DeepFake speech recordings.

Fine-tuned NLP networks to perform zero and few-shot author identification for hundreds of news articles.

Built calibrated classification frameworks to determine the text generator that created a piece of writing.

Radio Frequency Machine Learning Systems (DARPA RFMLS)

Designed and trained novel convolutional neural networks for the Radio Frequency (RF) domain.

Implemented parallel extraction of spectral features across GPUs to process 4 GB/s of input data.

Created a neural network gym training environment with an average throughput of 30 TB per day.

Key author of the multi-million dollar winning proposal, technical lead for day-to-day implementations.