Reconciliation Workbench — human-in-the-loop AI for tax compliance
Rules engine + LLM triage + HITL review + golden-dataset evals · multi-jurisdiction (US sales tax, EU CTC e-invoicing) · Aug 2026 — present
Human-in-the-loop AI exception review for multi-jurisdiction tax compliance. Deterministic core, AI at the edges, humans at the points of consequence: a rules engine detects six exception classes, an LLM triage layer explains each exception and proposes a disposition, a reviewer approves or overrides, and every decision lands in an exportable audit trail. An eval panel scores every run against seeded ground truth — 100% recall, zero false positives.
Measured on every run
Golden-dataset eval panel measures precision and recall against seeded ground truth on each run: 100% recall, 1.0 precision, zero false positives.
Jurisdictions as configuration
Adding a country is one JSON entry and zero downstream code changes. A declarative connector layer normalizes native-format feeds (semicolon CSVs, comma decimals, nested-JSON e-invoice statuses) into one canonical model.
Six exception classes
Rate, arithmetic, duplicate invoice, e-invoice linkage, orphan e-report, and return tie-out — detected deterministically, explained by the LLM, decided by a human.
Audit trail & governance
Every approve/override is logged to an exportable audit trail. All data synthetic; deployed auth-gated on Hugging Face Spaces; security-scanned with Bandit and pip-audit.
- Python
- Gradio
- Human-in-the-loop (HITL)
- Rules engine
- LLM triage
- Golden-dataset evals
- AI governance
- Synthetic data
- Hugging Face Spaces