SHOW HN: FinLang – A deterministic, auditable rules engine for finance (AGPL) (github.com)

FinLang is a new open-source (AGPL-3.0) domain-specific language and high-performance CLI engine for financial transaction categorization that emphasizes determinism, explainability and auditability — explicitly designed to satisfy EU AI Act “high‑risk” obligations. Instead of opaque ML classifiers, FinLang uses human‑readable .fin rules processed top-to-bottom (the last matching rule wins; flags accumulate append‑only), with a full audit trail (before/after diffs) and stateless execution for reproducibility. The project also ships a “Growth Loop” (Discover → Suggest → Categorize) that auto-discovers uncategorized patterns and generates candidate rules to accelerate coverage gains. Technically, FinLang is a vectorized engine built on Pandas + NumPy + PyArrow, claiming validated throughput up to ~39k rows/sec on a 100k UK dataset and ~24k rows/sec on a 5M×50‑col validation; audit modes add ~38% overhead. It supports locale-aware parsing, multi‑currency symbol stripping, CR/DR semantics and automatic amount synthesis across edge cases. The suggest flow reports 97.8% success on addressable patterns; exact-match emit is recommended for production (fuzzy tokenization has known limits). Install via pip (optional fastio extra for PyArrow). For AI/ML teams, FinLang is a practical alternative or complement to supervised models where determinism, traceability and regulatory compliance are paramount.