Data handling

This page covers the data you share for paid work. What the website collects from visitors is on the privacy page.

What data the work needs

FinDaS works on data you already hold, most often read straight from your data warehouse.

Nothing is installed on your systems. FinDaS works from its own side, as your team.

Queries against your warehouse only read: they select data and change nothing in it.

Where the query setup for your warehouse sets a rule for personal data, every query applies it.

The access key you issue for those queries is held only on FinDaS's working machine, outside version control and outside any backup.

Data can also arrive as files you export and send, by email or through a shared folder in FinDaS's Google Workspace.

No data is needed before the first call, and the booking page asks for none.

Where it is processed

FinDaS is based in Bulgaria, and the work runs on its own working machine.

Each client has a separate folder and a separate private repository on GitHub. One client's material never enters another's.

That repository holds the documents you send, the records of calls, and the code and reports FinDaS produces.

Row-level data, with one row per customer or account, never enters the repository. It stays in a folder on the working machine that is kept out of version control.

Files you send are sorted as they arrive: documents into the repository, row-level exports into that folder.

That folder is not backed up anywhere.

A retired working machine is wiped before it leaves FinDaS's hands, because it holds client data.

Finished work is posted to a client portal at clients.findas.org, behind one password per client. The portal runs on Cloudflare.

Drafts are shared for review through links that expire after 14 days by default, and FinDaS keeps them in its own task tracker while you review. Both also run on Cloudflare.

Which model providers it passes through

AI agents do the work and check it. Whatever data an agent reads is sent for processing to the model provider that runs it.

These model providers process client data today:

• Anthropic. Its Claude models run the agents that write the analysis and review it.
• OpenAI. Its Codex model is one of the three reviewers that check the code behind a deliverable.
• TypeSafe. Its Jev model classifies free-text fields and gives a first answer on the judgement calls inside a report. Names and other identifiers are stripped before text goes to it wherever the question does not need them. TypeSafe's data processing terms set no retention period.
• Google. Gemini's meeting notes in Google Meet summarise some calls.

Other call recordings are transcribed on the working machine itself, by a speech-to-text model that runs there, so the audio never leaves it.

The same providers work on every engagement.

Some of the services on this page process data outside the European Economic Area, including in the United States.

Model training is switched off on the Anthropic and OpenAI accounts FinDaS works through, so what their Claude and Codex agents are sent is not used to train those providers' models.

How long it is kept

FinDaS keeps your data for as long as the agreement runs.

FinDaS's standard service agreement binds it to keep the non-public information you share confidential for two years after the agreement ends.

Records of the engagement itself, such as the agreement and invoices, are kept for up to ten years after the work ends, in line with Bulgarian accounting and tax record-keeping.

When the work ends

Under the same agreement, FinDaS returns or destroys the confidential information you shared within 30 days of the agreement ending, or of your written request, and confirms in writing once it is done. That covers every copy, summary and extract, except the records the law requires FinDaS to keep, such as the agreement and invoices, and copies in backups or held by its service providers, which are deleted on their normal schedule.