LedgerMind

An AI-native ERP for accounting and financial control where every figure on a report drills back to the source document and the policy that produced it. The ledger cannot be wrong by construction: double entry is enforced by database constraints, money is held as scaled integers rather than floats, and posted journals are immutable. Accounting policy is written as executable code in a compiler-verified language, so a non-exhaustive branch or a currency mismatch fails to deploy instead of failing at quarter-end close. The audit trail is hash-chained and append-only, signed with Ed25519, so tampering is detectable rather than merely discouraged. Twenty-two segregation-of-duties rules are enforced at the capability level, before an action happens, instead of being reported afterwards. A hypothesis-based audit layer looks for the transactions that should exist and do not, which is the class of finding a conventional ERP never surfaces. Agents work inside the same governance: they can prepare, reconcile and explain, but the approval capability stays with a human. Multi-tenant isolation is enforced by Postgres row-level security, and complex tax cases such as UAE VAT reverse charge are modelled in policy rather than hardcoded into reports.

• Double entry enforced by database constraints, money as scaled integers, posted journals immutable
• Accounting policy as compiler-verified code: non-exhaustive branches and currency mismatches block deploy
• Every report figure drills through journal, invoice, purchase order and requisition to the source
• Hash-chained append-only audit log signed with Ed25519, so tampering is detectable
• Twenty-two segregation-of-duties rules enforced at the capability level, not reported after the fact
• Hypothesis-based audit that finds the transactions which should exist and do not
• Agents prepare and explain, humans approve: autonomous approval is not a capability agents hold
• Postgres row-level security for multi-tenant isolation, UAE VAT reverse charge modelled in policy