Skip to content
novelavia
novelavia
Try DemoGet Early Access
contact@novelavia.com

Technology

Document AI, validation and human review in one pipeline

Novelavia combines multimodal Gemini models with deterministic business rules and a human review step. The production architecture is designed for Google Cloud, starting lean and adding services as volume grows.

See it in the demo

Our AI approach

AI is the core of the product, not a chatbot on the side

Extraction quality decides whether automation saves time or creates rework. That is why every AI result passes through validation and, where needed, a person.

  • Layer 1

    Multimodal understanding

    Gemini models are designed to read the page as a whole, including layout, tables, stamps and handwriting, not just raw text. Output is constrained to a strict JSON schema.

  • Layer 2

    Layout-aware OCR where it helps

    Document AI is planned for low-quality scans and faxes, where dedicated OCR and layout parsing improve accuracy before reasoning.

  • Layer 3

    Deterministic validation

    Business rules check every result: required fields, date order, rate math, identifier formats, signatures and cross-document consistency.

  • Layer 4

    Human review

    Low-confidence, missing and inconsistent fields go to a reviewer. Every correction is recorded next to the original value.

Proposed architecture

An event-driven processing pipeline on Google Cloud

This is the planned production architecture. Today, the proof of concept runs extraction synchronously in a single request; asynchronous processing and storage arrive with the MVP.

  1. 01

    Document upload

    Web app or email ingestion

  2. 02

    Cloud Storage

    Private bucket, encrypted at rest

  3. 03

    Pub/Sub event

    Asynchronous processing

  4. 04

    Cloud Run service

    Processing workers

  5. 05

    Document AI + Gemini

    OCR, classification, extraction

  6. 06

    Field extraction

    Structured, typed values

  7. 07

    Business rule validation

    Deterministic checks

  8. 08

    Human review

    Uncertain fields only

  9. 09

    Database

    Firestore records

  10. 10

    API · JSON · CSV

    Exports and integrations

Secret Manager
Identity Platform
Cloud Logging and Monitoring
BigQuery analytics (later phase)

Planned platform services for the MVP and later phases. None of them are in production yet.

Google Cloud services

Start lean, scale deliberately

The MVP uses a small set of managed services: Cloud Run, Cloud Storage, Gemini and Firestore. Event-driven processing, dedicated OCR and analytics are added when real volume requires them.

PrototypeMVPLater phase

Prototype: implemented in the current proof of concept. MVP: part of the first production release. Later phase: introduced as volume and customer requirements grow. Nothing on this page is running in production yet.

  • Gemini on Gemini Enterprise Agent Platform

    Prototype

    Multimodal document understanding, contextual reasoning, field extraction and validation support. Formerly Vertex AI.

  • Cloud Run

    MVP

    Containerized web application, REST APIs and document processing services. This website is packaged for it.

  • Cloud Storage

    MVP

    Private, encrypted storage for uploaded documents and processing results with retention rules.

  • Firestore

    MVP

    Shipment records, document metadata, review state and processing status.

  • Secret Manager

    MVP

    Application credentials such as SMTP and integration secrets.

  • Identity Platform

    MVP

    Customer authentication and access management for the application.

  • Cloud Logging and Monitoring

    MVP

    Structured logs, metrics, alerting and latency tracking. Document contents are never logged.

  • Document AI

    Later phase

    OCR and layout-aware parsing for low-quality scans where it improves accuracy.

  • Pub/Sub

    Later phase

    Asynchronous, event-driven processing for batches and email ingestion.

  • BigQuery

    Later phase

    Document analytics, accuracy evaluation datasets and operational insights.

  • Cloud SQL

    Later phase

    Relational option for accounting-grade records if reporting needs outgrow Firestore.

  • Cloud KMS

    Later phase

    Customer-managed encryption keys for enterprise customers who require them.

In the proof of concept today

How live analysis works

The demo includes a server-side extraction endpoint that calls a Gemini model through Google Cloud. It is switched off on public deployments unless explicitly enabled.

  1. 1The browser sends the file to Novelavia's own API. It never talks to an AI service directly.
  2. 2The server checks the file signature, type and size, and applies rate limits.
  3. 3The document is sent to a Gemini model as inline data, together with strict extraction instructions and a JSON schema.
  4. 4The response is normalized: unknown fields are dropped, missing required fields are added as empty, never guessed.
  5. 5Validation rules run on the result, and the review interface highlights what needs attention.
  6. 6The file is processed in memory for that single request and is not stored. Logs contain metadata only.

Security and data privacy

Freight documents are confidential. We treat them that way.

Rate confirmations and invoices contain pricing, customer relationships and sometimes personal information. These principles shape the architecture from the start.

In the proof of concept today

  • Uploads checked by file signature, type and size before processing
  • Rate limits on public endpoints
  • No logging of document contents or extracted values
  • AI models are called only from the server; no API keys are exposed
  • Security headers such as HSTS and frame protection
  • Human verification of uncertain AI output in the review interface

Planned for the MVP

  • Private storage with no public access to uploaded documents
  • Encryption at rest with Google-managed keys, and customer-managed keys for enterprise customers
  • Configurable document retention and deletion
  • Least-privilege service accounts, with credentials kept in Secret Manager

Protection against document-based prompt injection

  • Document content is passed to the model strictly as data, with instructions that it must never be followed
  • Responses are constrained to a fixed JSON schema; free-form output is rejected
  • Server-side normalization drops unknown fields, clamps values and limits lengths
  • The model has no tools or actions available, so text in a document has no operation to trigger
  • Text that tries to address an AI system is reported to the reviewer as a warning

Novelavia does not currently hold SOC 2 or any other compliance certification. We will not claim one until an independent audit has been completed.

Measuring quality

How we will evaluate accuracy

We benchmark extraction on labeled, anonymized documents before making any performance claims. No accuracy figures are published yet.

  • Field-level accuracy

    Exact and normalized match per field and per document type.

  • Confidence calibration

    Whether low-confidence fields are the ones that are actually wrong.

  • Latency

    Median and 95th percentile processing time per page.

  • Cost per document

    Model and infrastructure cost, tracked per document type.

Help shape Novelavia from day one.

We are looking for U.S. carriers, brokers, dispatch services and freight accounting teams to pilot the MVP. Early access members try new capabilities first and directly influence the roadmap.

Get Early Access Try Interactive Demo