Radix-Research

Welcome to the first entry of my Radix research journal. This blog will document the progress of my MSc Computer Science with Data Science project, bridging the gap between academic data science and commercial enterprise engineering.

Within the B2B fintech sector, institutions rely heavily on processing statutory corporate data from Companies House. However, a massive portion of this historical intelligence is locked inside unstructured, scanned image files. When the industry attempts to automate the extraction of these documents, the pipelines frequently break down.

The B2B Financial Data Bottleneck

My research sits at the intersection of Financial Data Automation and Natural Language Processing (NLP), directly addressing the failures of current extraction methods:

• The Collapse of Spatial Geometry
  Standard Optical Character Recognition (OCR) relies on spatial bounding boxes. If a scanned financial statement is skewed by even a few degrees, the spatial geometry collapses, causing separate accounting columns to bleed into a single, fractured text string. As Kim et al. (2022) note, traditional OCR frequently introduces compounding, unrecoverable errors into document pipelines.

• Manual Interventions
  To circumvent OCR failures, the industry often reverts to expensive and slow manual data entry (FIS Global, n.d.).

• Data Sovereignty Risks
  Routing these massive, unredacted corporate documents through external, commercial Large Language Models (LLMs) creates severe data privacy risks, violates enterprise data sovereignty frameworks (Deloitte, 2024), and incurs unsustainable token costs.

The Proposed Solution: Zero-Touch Financial Spreading

To resolve this bottleneck, I am evaluating whether a locally hosted, domain-specific Small Language Model (SLM) can act as an intelligent neural autocorrect. This project proposes an air-gapped, two-stage software architecture:

1. High-Fidelity Optical Extraction: Extracting raw text from degraded Companies House PDFs using modern OCR-free Vision Transformers (ViTs) to output structurally intact Markdown. This completely avoids hardcoded text-cleaning rules (Regex), which often destroy financial layouts.

2. SLM Semantic Normalisation: Passing this visually noisy Markdown into a fine-tuned SLM, operating locally on an RTX 3060 via QLoRA quantisation. The model will dynamically map the fractured strings to standard FRS 101/102 JSON formats based on semantic context rather than brittle spatial coordinates.

Evaluation and Outcomes

The ultimate goal is to prove that by fine-tuning an SLM on verified iXBRL ground truth, we can achieve the same taxonomy standardisation accuracy as proprietary APIs like GPT-4o. Accuracy will be rigorously evaluated using Exact Match (EM) for key-value pairs and F1 scores for semantic similarity.

By running this entirely offline on consumer-grade hardware, the final deliverable is a standalone Node.js application with a transparent processing log, aims to deliver reduced and predictable inference latency with absolutely zero data compliance risks.

Over the coming weeks, I will be documenting the system implementation, testing phases, and performance metrics right here.

References

• Deloitte (2024) Navigating Generative AI in Financial Services: Security, Privacy, and Sovereign Frameworks. [Industry White Paper].

• FIS Global (n.d.) Automated Financial Spreading (AFS). [White Paper].

• Kim, G., Hong, T., Yim, M., Nam, J., Park, J., Yim, J., Hwang, W., Yun, S., Han, D. and Park, S. (2022) 'OCR-free Document Understanding Transformer', European Conference on Computer Vision.