Case study · AI & Automation
Taxonaut: accounting automation for Polish sole traders
Self-hosted auto-accountant for Polish JDG: invoices in, tax declarations out, with rules-as-data and tiered LLM processing.
- Industry
- Accounting & tax
- Timeline
- 8 weeks, ongoing
- Year
- 2026
- Services
- Product engineering, AI integration, Compliance research
The problem
A Polish sole trader (JDG) burns hours every month on the same loop: collect invoices, classify costs, compute ZUS and PIT advances, fill government forms, try not to miss a deadline. Accountants charge monthly retainers for what is, most months, deterministic paperwork. Off-the-shelf SaaS tools automate fragments, hold your data hostage, and still make you re-type numbers into government portals.
What we built
A self-hosted automation platform that treats accounting as a pipeline: documents and bank statements go in, ready-to-submit declarations and payment reminders come out.
Three design decisions carry the whole system:
- Rules as data, not code. Polish tax rules change yearly. Rates, thresholds, form versions and computation rules live in versioned database tables with effective dates, so a rule change is a data migration, not a redeploy. The 2026 rates coexist with 2025 history cleanly.
- Submit-as-is documents. Every government-bound artifact the system generates is validated against the current official schema. The hard rule: if it cannot be submitted without manual re-typing, it is a bug.
- Tiered LLM processing. Bulk work like invoice field extraction runs on cheap fast models through a gateway. Critical-tier decisions, like ambiguous cost classification with tax consequences, are queued for a frontier model and processed asynchronously through a job queue, keeping quality where it matters and costs where they belong.
Architecture
FastAPI plus PostgreSQL (multi-tenant with row-level isolation), Celery workers and a scheduler for the monthly cycle, a Telegram bot for capture and reminders: photograph an invoice, get it parsed, classified and filed. A reminder service tracks statutory deadlines and nags with escalating urgency.
Why it matters
This is the pattern for a whole class of businesses: regulated, repetitive, deadline-driven paperwork where correctness is non-negotiable. Rules-as-data plus schema-validated outputs plus tiered AI is a template we can lift into any compliance-heavy domain.