Hours
Weeks
Commits
Role
Saldo
AI-Orchestrated Financial Automation with Self-Healing Imports and Government-Grade Compliance
Saldo is an AI-powered financial automation platform that turns messy real-world statements, invoices, and receipts into reliable, compliant bookkeeping data.
saldo-so/saldo
Saldo is an AI-powered financial automation platform that turns messy real-world statements, invoices, and receipts into reliable, compliant bookkeeping data. It targets small and mid-sized businesses that drown in manual reconciliation and government e-invoicing rules, especially in markets like Serbia with SEF and UBL requirements. Saldo ingests bank exports, camera photos, PDFs, and e-invoices, then uses AI to normalize, match, and route them so teams spend time approving financial actions instead of keying data. Under the hood, Saldo separates AI reasoning from execution. A two-phase tool orchestration layer has LLMs propose actions (e.g., create invoice, match bill to transaction) as structured previews requiring human confirmation; a hardened executor then applies them with Prisma-backed validation, encryption, audit logging, and cache revalidation in a Next.js/TypeScript stack. Bank statement parsing is modeled as declarative configs that are generated and self-healed by AI when formats change, while a custom vision-augmented parser combines PDF text extraction, image understanding, and Serbian tax rules to structure invoices and receipts. Government e-invoicing is abstracted behind a common connector interface that handles UBL generation, SEF/Peppol submission, and async status polling without leaking regulatory complexity into the rest of the app. To support real usage, Saldo includes a hybrid transaction-to-document matcher that blends rule-based scoring with AI-enhanced vendor extraction, a document scanning pipeline that performs perspective correction for mobile photos, and subscription-aware middleware that gates AI features by Polar plans and per-organization quotas. This turns AI-heavy workflows into predictable, auditable operations suitable for production finance teams rather than demo-level prototypes.
Engineered an AI-driven, self-healing bank statement importer that regenerates and validates per-bank parsing configs, dramatically reducing manual mapping work as statement formats change over time.
Built a vision-augmented financial document parser that combines PDF text extraction, image understanding, and Serbian-specific tax rules to reliably structure invoices, receipts, and bills from noisy real-world uploads.
Designed a two-phase AI tool orchestration layer where LLMs propose financial actions for human confirmation, then a hardened executor applies them with encryption, audits, and cache revalidation.
Developed a hybrid transaction-to-document matching engine that blends rule-based scoring with AI-enhanced vendor and entity extraction to accurately reconcile bank transactions against uploaded bills and receipts.
Implemented pluggable government e-invoicing connectors that validate, generate UBL, submit, and asynchronously poll statuses for SEF and Peppol, isolating regulatory complexity behind a unified interface.
Orchestrated an invoice sending pipeline that renders compliant PDFs, dispatches email, and conditionally submits to SEF with pre-checks and idempotent status transitions across draft, sent, accepted, and rejected states.
Created a custom document scanning pipeline that detects corners, applies perspective correction, and resamples images to produce scanner-quality receipts and bills from imperfect camera photos.
Established subscription-aware middleware that reads Polar plans to gate AI features, enforce per-organization quotas, and return traceable errors, enabling predictable, usage-based monetization of automation features.
Saldo automates end-to-end financial paperwork for SMEs by extracting, reconciling, and submitting documents with AI, reducing manual bookkeeping and compliance overhead across banks, invoices, and government systems.