Skip to main content

AI Governance & Data Privacy

Zero Training on Customer Data

A binding contractual and architectural commitment: Dotnitron never uses customer confidential data, diligence workpapers, prompts, or completions to train public or foundation AI models.

1. Our Unconditional Commitment

Contractual Guarantee

Dotnitron Technologies Private Limited contractually and architecturally ensures that no client confidential files, transaction documents, financial models, reviewer notes, prompts, model completions, or vector embeddings are ever used to train, retrain, fine-tune, or improve any public, foundational, or shared AI models. All model inference across our products and client engagements is executed exclusively through enterprise commercial endpoints bound by verified zero-training terms, or directly within customer-controlled private infrastructure.

This commitment is enshrined in our standard Data Processing Addendum (DPA) and individual enterprise Master Services Agreements (MSAs).

2. Architectural Enforcement by Deployment Model

Because Dotnitron delivers both SaaS products and forward-deployed engineering solutions for banking, private equity, and advisory firms, we enforce zero-training through two distinct delivery models:

Model A: Enterprise Engagements

Client VPC / On-Premise (BYOK / BYOA)

The software is deployed directly inside the client's cloud boundary (AWS VPC, Azure Tenant, GCP Project) or private on-premise hardware.

  • Client provides their own model API keys (BYOK) under their existing enterprise agreements.
  • Client VPC firewall rules prevent external telemetry or outbound egress.
  • Supports fully air-gapped, local model execution (e.g., Llama 3 / vLLM on private GPUs).
  • Geographic data residency is 100% controlled by the client.
Model B: Hosted Products & Pilots

Dotnitron-Managed Cloud

Hosted deployments utilize single-tenant workspaces or dedicated database schemas in customer-selected cloud regions (e.g. India, US, EU, UK).

  • Model inference is routed strictly through verified commercial API endpoints with zero-training terms.
  • Optional BYOK toggle enables clients to provide their own OpenAI or Azure credentials.
  • Encrypted at rest (AES-256) and in transit (TLS 1.3).
  • Tenant-level vector isolation in Qdrant ensures zero cross-customer data leakage.

3. Provider-by-Provider Contractual Verification Matrix

Below is the comprehensive legal and technical audit of all AI model providers and supporting infrastructure used across Dotnitron systems:

ProviderEndpointsZero-Training PolicyRetention & LoggingBYOK Support
Microsoft Azure OpenAIMicrosoft Customer Agreement, Cognitive Services Terms & DPAGPT-4o, GPT-4o-mini, o1, custom Azure deploymentsVerified

Zero training. Prompts, completions, embeddings, and fine-tuning data are never used to train or improve any Microsoft or third-party models.

Ephemeral inference processing. Standard 30-day abuse monitoring can be modified or exempted with approved Zero Data Retention (ZDR).Yes (Client Azure Subscription / Tenant) or Dotnitron endpoint
OpenAI API (Enterprise / Commercial)OpenAI Business Terms & Commercial API Data Privacy Policygpt-4o, gpt-4o-mini, text-embedding-3-small/largeVerified

Zero training. Data submitted to and generated by the commercial API is not used to train or improve OpenAI models (in effect since March 1, 2023).

Stored for up to 30 days solely for abuse/misuse detection, then deleted. Zero Data Retention (ZDR) available for qualifying enterprise accounts.Yes (Client API Key) or Dotnitron endpoint
Anthropic Commercial APIAnthropic Commercial Terms of ServiceClaude 3.5 Sonnet, Claude 3.5 Haiku, Claude 3 OpusVerified

Zero training. Anthropic does not train generative foundation models on customer data, prompts, or completions submitted through the commercial API.

Ephemeral inference processing. Prompt caching is restricted to the customer organization workspace.Yes (Client Anthropic Key) or Dotnitron endpoint
Google Cloud Vertex AIGoogle Cloud Master Agreement & Service Specific TermsGemini 1.5 Pro, Gemini 1.5 Flash, Multimodal endpointsVerified

Zero training. Customer data (prompts, completions, embeddings) is not used to train Google foundation models without customer consent.

Data stays encrypted in customer-selected regions; supports Customer-Managed Encryption Keys (CMEK).Yes (Client GCP Project) or Dotnitron endpoint
AWS BedrockAWS Service Terms & AWS GDPR Data Processing AddendumClaude, Llama, Titan, Mistral via AWS APIVerified

Zero training. AWS does not use customer inputs or model outputs to train any base foundation models or share data with model providers.

All data remains strictly within the customer's selected AWS Virtual Private Cloud (VPC) and region.Yes (Client AWS Account / Role)
Azure Document IntelligenceMicrosoft Cognitive Services Product Termsprebuilt-layout, OCR, table extractionVerified

Zero training. Source PDFs and scans are processed ephemerally for structure extraction and never retained to train models.

Ephemeral processing; deleted immediately after structured layout extraction is returned.Yes (Client Azure Resource) or Dotnitron endpoint
AssemblyAIAssemblyAI Paid Commercial API Termsuniversal-2 speech-to-textVerified

Zero training. Audio files, transcripts, and speaker diarization are not used to train base speech models.

Configured for automated deletion following completed transcript delivery.Yes (Client Key) or Dotnitron endpoint
Local & Private Models (Air-Gapped)Open Source Permissive Licenses / Self-HostedLlama 3, DeepSeek, Mistral (vLLM / Firecracker microVMs)Verified

Absolute zero egress. Models run on dedicated customer GPUs or isolated on-premise servers with no internet connectivity required.

100% customer-controlled. Weights are static; no data leaves the physical or VPC host.Customer-hosted infrastructure
Qdrant Vector DatabaseQdrant Cloud DPA / Apache 2.0 Self-HostedSemantic retrieval, vector chunk indexingVerified

Zero training. Dedicated vector storage engine; embeddings are stored in customer-isolated collections, never shared with LLMs.

Deleted automatically upon document or workspace purge.Yes (Self-hosted or Managed)

4. Human-in-the-Loop & AI Safety Governance

Protecting customer data is paired with our rigorous AI governance principles:

  • Mandatory Citation Grounding: AI-generated assertions in diligence memos and workpapers must explicitly reference underlying source documents with page-level or clause-level citations.
  • Human-in-the-Loop Review: AI outputs are treated as assistive drafts. Final investment decisions, regulatory filings, and credit approvals require affirmative human professional sign-off.
  • Prompt Injection Defenses: All incoming document chunks and user prompts are passed through untrusted content delimiters, input sanitization routines, and refusal gates before model ingestion.

For detailed operational standards, see our published AI Usage & Acceptable Use Policy.