Privacy-firstOpen-source modelsNo training on your data
Private by design. Smarter by default.
We use open‑source AI to automatically categorize your transactions — but only after we anonymize them. Your raw banking data never leaves our system, and it’s never used to train any model.
How it works
Categorization in three secure steps.
1. Anonymize
We hash merchant information, sanitize descriptions, and include only what’s needed: amount magnitude and whether it’s a credit or a debit. No personally identifiable information is sent to AI.
2. Sign & send securely
Requests are HMAC‑signed and time‑bound (5‑minute window) before they reach our serverless worker. We do not log raw transaction payloads.
3. Classify with open‑source AI
Categorization runs via dedicated infrastructure using open‑source models such as Gemma or Mistral. Model providers do not use your data for training.
Exactly what we send
{
merchantHash: string,
descriptionSanitized: string,
amount: number, // absolute amount (no currency, no account info)
transactionType: "credit" | "debit"
}That’s it — no names, no account numbers, no raw merchant details.
What we never send
- Account identifiers (IBAN, card numbers, user IDs)
- Raw merchant names or untouched descriptions
- Personal details from your profile
- Any data for model training
Privacy & AI FAQ
Short answers to the most important questions.
Why open‑source models?
Transparency and control. We run models such as Gemma or Mistral via dedicated infrastructure, so your data stays out of vendor training pipelines and we can audit prompts and outputs.
Is my data used to train AI?
No. We never use your data to train any model. Requests are anonymized, signed, and processed for categorization only.
Can I review or change categories?
Yes. You can always review and edit categorizations. When the model is unsure, we flag items for a quick human review.