Contabooks S.R.L. · Moldova
Product & technology
ContaBooks is a subscription software platform for small and medium-sized businesses in Moldova. It connects invoicing, accounting, inventory, payroll and financial document workflows in one multilingual workspace.
The product we are building
We develop our own SaaS product for business owners, accountants and finance teams. ContaBooks brings customer records, invoices, contracts, bank statement imports, double-entry accounting, inventory and payroll into a shared company workspace. Reporting tools prepare TVA12, IPC21 and IRM19 drafts and versioned XML; official submission and acceptance depend on the relevant government service and the company’s configuration.
From documents to financial visibility
A team creates a customer and invoice, shares the document, imports a bank statement and reconciles the payment with its source document. Accounting entries and reports help the team review financial activity. AI tools assist with receipt extraction, expense categorisation and document review. Users review suggestions and remain responsible for accounting decisions and external submissions.
Customers & business model
ContaBooks is a product subscription business serving Moldovan small and medium-sized companies and their finance teams. A free entry plan supports basic invoicing; paid monthly plans expand access to accounting, inventory, payroll, reporting, AI and team features. The public pricing page describes the current plans. Customers pay for access to the software.
Application architecture
The web application uses Next.js, React and TypeScript, with Romanian, Russian and English interfaces. A NestJS API implements business workflows, backed by PostgreSQL and Prisma. Company-scoped access checks and roles control access to business records. API integrations connect document, banking and reporting workflows with external services where configured.
AI & automation
The backend implements Gemini through Google Cloud Vertex AI with the Google Gen AI SDK. Text and image workflows support document extraction, classification, summaries and review suggestions. A dated direct model-call example is published below. Each application workflow still depends on its server credentials, access checks and output validation.
Why AI matters to ContaBooks
Receipts, invoices, contracts and reporting errors arrive as unstructured text or images. AI is a core automation layer that turns this material into draft fields, classification suggestions and plain-language explanations in Romanian, Russian and English. It reduces repetitive interpretation before a user reviews the result. Accounting arithmetic, tax formulas, XML generation and official filing states remain controlled by deterministic application services.
Access & data protection
The application implements authenticated access, company-level roles and data access checks. Selected integration secrets are encrypted at application level. Bank-hosted checkout keeps card entry on the payment provider’s page. These are application controls, not a claim of independent security certification. Privacy and contractual information are available through the links below.
Google Cloud: present capability and next steps
The code includes Gemini/Vertex AI integration, and the owner recorded a successful direct model call from Cloud Shell on 25 October 2026. This verifies the project and caller used for that test, not the deployed application server identity or full hosting architecture. The roadmap below connects proposed cloud services to product workloads; migration remains subject to validation.
How AI fits into a business workflow
01
Start with a document or authorised context
A user supplies receipt text or an image, or opens a review using company-scoped document and reporting context. Only the context needed by that workflow should be sent to the model.
02
Request a focused Gemini response
The backend sends a task-specific prompt through Vertex AI. Extraction asks for structured fields; review and reporting assistants ask for findings or explanations.
03
Check the returned data
Application code parses and validates the output for the selected task. Reporting copilots bind explanations to source versions and reject stale context. A model response is a proposal, not a ledger entry or official acceptance.
04
Keep the user in control
The user reviews extracted values and suggestions. Saving, recalculation and supported approval actions use the existing domain commands and permission checks.
REAL MODEL RESPONSE · SYNTHETIC INPUT
A recorded Vertex AI extraction example
On 25 October 2026, the project owner ran a direct Gemini request from Google Cloud Shell. The synthetic invoice text and returned JSON below come from the supplied execution transcript; this is separate from the prepared AI examples in the product screenshots.
- Returned model version
- gemini-2.5-flash
- Provider and region
- Vertex AI · us-central1
- Provider response timestamp
Synthetic input sent to the model
Synthetic demo invoice: total 2500 MDL, date 2026-10-25. Extract amount, currency and date as JSON.
JSON returned by the model
{
"amount": "2500",
"currency": "MDL",
"date": "2026-10-25"
}The returned amount, currency and date match the synthetic input. The response completed with STOP and reported 142 total tokens. This is one test case, not an accuracy or performance benchmark.
Scope: an owner-run Cloud Shell API call using the owner’s Google identity. The transcript was supplied by the owner and has not been independently replayed. It does not verify the deployed ContaBooks server credentials, a full authenticated application workflow or accounting correctness.
View the request, response and provenance recordGoogle Cloud workload plan
Credits would support AI workload evaluation first, followed by measured infrastructure pilots. The statuses below separate demonstrated capability from proposed architecture; they are not a migration completion claim.
| Service | Planned workload and validation | Current evidence / stage |
|---|---|---|
| Vertex AI / Gemini | Document understanding, structured extraction and contextual explanations. Evaluate a synthetic test corpus, response validation, usage and cost before widening access. | Client implemented; direct Cloud Shell model call recorded on 2026-10-25. Application-server identity check remains open. |
| Cloud Run | Evaluate container hosting for the API and web application, including request latency, scaling limits and compatibility with background tasks. | Proposed pilot; hosting migration is not confirmed. |
| Cloud SQL for PostgreSQL | Evaluate managed PostgreSQL using a non-production copy, migration compatibility checks and a backup/restore drill. | Proposed pilot; current database migration is not confirmed. |
| Cloud Storage | Evaluate private storage for documents and generated exports, with access rules, retention and deletion tested before customer use. | Proposed design; no completed storage migration claimed. |
| Secret Manager / IAM | Evaluate workload identities, narrowly scoped service permissions and managed secret access for deployed workloads. | Cloud Shell used the owner’s identity. Production workload identity and secret migration require separate verification. |
| Cloud Monitoring | Measure request failures, latency, model usage and spend; set budgets and alerts before increasing the pilot workload. | Proposed operational acceptance criteria; no measured cloud baseline published yet. |
Pilot acceptance should cover a repeatable synthetic evaluation set, task-specific output checks, company access isolation, user review, failure handling, restore checks and budget alerts. Funding stage, credit eligibility and actual cloud spend are stated separately in the application.