04 / FINTECH

SALDO

Turn messy bank exports, photos, and PDFs into compliant ledger data. Meet Serbia's SEF and UBL requirements without manual reconciliation.

OUTCOME: Two-phase orchestration — AI proposes, hardened executor applies. Self-healing import configs, vision-augmented parsing, pluggable e-invoicing connectors.
STACK: Next.js · Vercel AI · Postgres · Redis

AI proposes, a hardened executor applies: Financial data is the wrong place to let a model hold the write path. The pipeline splits in two: the model produces a structured proposal describing what it believes each row means, and a separate executor validates it against a schema, checks tenancy and invariants, and applies it inside a transaction. A rejected proposal is a normal outcome, not an error.
Imports that repair themselves: Bank exports change format without warning, and a hardcoded parser breaks silently on the next statement. Import configurations adapt to the shape of what actually arrived, with vision-augmented parsing handling photographed and PDF documents that never had a clean column structure to begin with.
Compliance as a connector: UBL, SEF, and PEPPOL are moving targets with jurisdiction-specific rules. Treating e-invoicing as a pluggable connector rather than a core assumption means a regulatory change is a connector change — not a migration through the ledger.

CAPABILITIES SHIPPED HERE

HUMAN-IN-THE-LOOP: AI proposes as structured previews. A hardened executor applies — with validation, encryption, and audit trails throughout.
MULTI-MODEL & BILLING: OpenAI, Anthropic, Google under one orchestration. Token-to-credit ledgers per organization. Attribution across every call.