Ertas for Compliance Report Generation

The Challenge

Compliance reporting is one of the most labor-intensive functions in regulated industries. Financial institutions file quarterly reports with the SEC, OCC, and FDIC. Healthcare organizations submit quality measures to CMS and state health departments. Environmental companies report emissions data to the EPA. Each regulatory body demands specific formats, terminology, calculation methodologies, and disclosure requirements, and the penalties for errors or omissions range from fines to criminal liability.

Compliance teams spend hundreds of hours per reporting cycle gathering data from multiple systems, formatting it according to regulatory specifications, drafting narrative sections in the precise language regulators expect, and reviewing every number and statement for accuracy. Generic AI writing tools are dangerously unsuitable for this work. They generate plausible-sounding text that uses the wrong regulatory terminology, omits required disclosures, or makes factual claims that do not align with the underlying data. In compliance, approximately correct is not correct at all.

The Solution

Ertas lets compliance teams train a custom model on their organization's historical compliance reports, regulatory filing templates, and examiner feedback. The model learns the regulatory language, required disclosures, formatting conventions, and data presentation standards for each filing type, so its drafts follow the patterns of filings that were already approved.

In Ertas Studio, the team uploads pairs of source data and the matching approved report sections as a JSONL dataset, or builds the pairs in Data Craft, Studio's guided dataset builder, where compliance staff edit, version and approve each example. Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas: cut it down to what the model needs and remove customer and personal data first. Once a run passes the team's own checks, the model is downloaded from Hub as GGUF or safetensors and run on the organization's own servers, so the data behind each new report and the drafts the model writes stay in-house. The model produces first-pass sections for human review, and approval and sign-off stay in the team's existing review process.

Key Features

Example Workflow

A mid-size bank's compliance team prepares quarterly Call Reports for the FDIC and annual CCAR stress test submissions for the Federal Reserve. The team collects five years of approved filings, 40 quarterly reports and 5 annual submissions, and pairs each narrative section with a summary of the structured data behind it. Customer-level detail is removed before the dataset is uploaded to Ertas Studio, and compliance analysts review the pairs in Data Craft, approving the ones that reflect current reporting practice. The team runs fine-tuning on Qwen 2.5 7B and Llama 3.1 8B in parallel and compares the drafts each produces against a held-out set of recent filings. The better model is exported as safetensors and served with vLLM on the bank's own servers, with access limited to the compliance team. In the next reporting cycle, the model drafts first-pass narrative sections from that quarter's data, and senior compliance officers review, edit and approve every section through the bank's existing sign-off process before anything is filed.