A high-precision RAG system designed to bridge the “Hallucination Gap” in technical document audit.

📊 Project: LLM Spreads (Proof of Concept)

The Inspiration: Solving the “70/30” Problem

During my time as a Credit Analyst in the banking industry, I experienced a frustrating reality: roughly 70% of my time was spent on the manual transcription of financial data from PDFs into spreading software. This left only 30% of my time for the actual analysis—the part of the job that adds real value to the company and requires human judgment.

I built LLM Spreads as a proof of concept to solve this “70/30” problem. The goal is to automate the tedious data entry phase, allowing analysts to focus on what they do best: evaluating risk and identifying opportunities.

Project Scope & Transparency

⚠️ Disclaimer: Prototype Version This project is a small-scale demonstration of how Generative AI can be applied to financial workflows. While the pipeline is fully operational, it was fine-tuned on a limited dataset of 29 high-quality examples.
In an institutional setting, this would require a dataset of 500+ documents to handle all possible reporting edge cases. This project serves as a showcase of my ability to design agentic workflows, fine-tune Large Language Models (LLMs) on Vertex AI, and deploy serverless AI applications.

Technical Highlights

graph TD Start([User Uploads PDF]) --> Parsing[Docling Parsing] Parsing --> Security{🛡️ Security Agent} Security -- "Unsafe" --> Alert([Stop & Alert]) Security -- "Safe" --> Financial[🔍 Financial Agent] Financial --> Verification[🧐 Verification Agent] Verification --> Output([✅ Structured JSON Output]) style Start fill:#f9f,stroke:#333,stroke-width:2px style Parsing fill:#e1f5fe,stroke:#01579b style Security fill:#ffcdd2,stroke:#c62828 style Alert fill:#b71c1c,stroke:#333,stroke-width:2px,color:#fff style Financial fill:#e8f5e9,stroke:#2e7d32 style Verification fill:#fff3e0,stroke:#ef6c00 style Output fill:#bbf,stroke:#333,stroke-width:2px

Custom Fine-Tuning: I performed Supervised Fine-Tuning (SFT) on Gemini 2.0 Flash via Vertex AI to ensure the model adheres to strict financial schemas and specific accounting formatting.
Agentic Verification Architecture: The app doesn’t just “guess.” It uses a linear multi-agent flow:
1. 🛡️ Security Agent: Scans for prompt injections or malicious input.
2. 🔍 Financial Agent: Performs the core extraction using the tuned model.
3. 🧐 Verification Agent: Acts as a “Senior Editor,” checking the extraction against the source text and correcting errors in real-time.
High-Fidelity Parsing: Utilizes the Docling library to accurately parse complex, multi-column financial tables into clean Markdown before processing.
Serverless Deployment: The final application is built with Streamlit and deployed via Modal, showcasing a scalable, production-ready deployment strategy.

The Results

Even with a small training set, the tuned model (spreads_llm2) showed measurable improvements in output formatting and style consistency compared to the base model.

Metric	Base Gemini 2.0	Tuned Model (`spreads_llm2`)
Schema Adherence	10.00	10.00
Accounting Logic	9.80	9.80
Format & Style	9.64	9.68

What’s Next?

This demo lays the groundwork for more complex financial automation, including:

Business Tax Returns: Expanding the schema to handle IRS Forms 1120 and 1065.
Scaling the Lab: Increasing the training set to 500+ examples for institutional robustness.
Granular Spreading: Automating more complex items like EBITDA adjustments and debt schedules.

LLM Spreads

End-to-End Financial Data Extraction!