PipePilot AI data and security statement

Last checked: 12 October 2026. Every provider claim below was read on the provider's own page on that date and is quoted with its address. Providers change their terms; if a page now says something different, the page wins and this statement is out of date.

This statement covers the PipePilot Assistant (Report > Assistant). The rest of PipePilot (building, running and checking a network, saving projects) runs in your browser and sends nothing to any AI provider. It is published at geopilotai.com/pipepilot/ai-data and kept in the source repository as pm/pipepilot/ai-data-and-security.md; the two are the same file.

In short

The assistant sends your question and a short summary of the open model, never the .inp file.
You choose where it goes, in this order: a model on your own computer, then your own provider key if you add one, then GeoPilot's India region (the default), then other cloud providers only if you allow them.
DeepSeek, whose servers are in China, is off by default. It cannot be used while the India region is ticked, and we recommend that government users never turn it on.
Each answer is labelled with the tier that gave it.
PipePilot results support professional judgement and do not replace it (see the last section).

What the assistant sends

Each time you ask, your browser builds a request containing:

Your question, and the last four questions of the conversation with the assistant's replies.
A model summary of at most 4,000 characters: the file name and title (first 80 characters each), object counts, units, the analysis options that change answers, the run status, the lowest pressures and highest velocities with their object IDs, the CPHEEO compliance counts, the model check counts, and the currently selected object's ID.
Tool results. When the assistant looks something up, your browser runs the lookup against the open model and sends back the result: for example the properties and results of one junction or pipe, or a list of up to 100 object IDs with their values.

What it never sends

The .inp file or the whole network.
Map coordinates, background maps or imported GIS layers.
Object descriptions and tags.
Your saved projects, versions or scenarios.

Note that the file name and model title are sent. If they identify a place or a client, rename the file before asking.

What GeoPilot keeps

Local model tier: nothing. Your browser talks to your own server; GeoPilot's server is not involved.
Hosted tiers (your own key, India region, other cloud): for every model call, GeoPilot writes one usage record to its own database: your account email, the first 500 characters of your latest question, which model answered, and token counts. It is used for cost accounting and abuse control and is kept with your account data under the Privacy Policy. That database is hosted by Supabase in Sydney, Australia (AWS ap-southeast-2), not in India. So even when the India region answers, your email and the start of your question are stored in Australia. The model summary and tool results are not stored.
Your own key is never written to GeoPilot's database or logs. It travels with each question over HTTPS, is used to make that one call, and is discarded. In the browser it is kept in memory for the open tab unless you tick "Remember the key on this device", which saves it in this browser's storage for PipePilot on this device.
Settings (which tiers are on, your local server address) are saved in this browser only.

The tiers and how to choose them

Open the Assistant, then Settings. Tiers are tried in the order below; a tier is used only when it is ticked. If a tier fails, the next ticked tier is tried. The label under each answer says which tier gave it.

1. Local model (on this computer)

Your browser sends the request straight to an OpenAI-compatible server on your machine, such as Ollama or LM Studio. Nothing goes to GeoPilot or any cloud provider. The tool protocol and prompt are the same as the hosted tiers, so the assistant can run the same lookups. Choose a model that supports tool calling.

Set up:

Start the server. Ollama serves at http://localhost:11434/v1; LM Studio at http://localhost:1234/v1.
Allow PipePilot's web address to call it (CORS):
- Ollama: set the environment variable OLLAMA_ORIGINS=https://geopilotai.com and restart Ollama. Ollama's FAQ: "Ollama allows cross-origin requests from 127.0.0.1 and 0.0.0.0 by default. Additional origins can be configured with OLLAMA_ORIGINS." (docs.ollama.com/faq)
- LM Studio: in the server settings, switch on "Enable CORS", or start the server with lms server start --cors. LM Studio: "Enable CORS : Switch Enable Cross-Origin Resource Sharing (CORS) to allow applications from different origins to access the API." (lmstudio.ai/docs/developer/core/server/settings). LM Studio warns that CORS is a security risk and recommends turning on authentication; if you do, put the key in PipePilot's optional Key box.
In PipePilot, tick Local model, enter the server address and the model name, and press Test connection. It lists the server's models and says whether yours is among them.
Chrome 142 and later asks permission the first time a website calls a server on your own machine. Allow it for geopilotai.com. Chrome: "The Local Network Access permission prompt is launching in Chrome 142." (developer.chrome.com/blog/local-network-access)

Ollama documents tool calling on /v1/chat/completions (docs.ollama.com/api/openai-compatibility); so does LM Studio (lmstudio.ai/docs/developer/openai-compat/tools).

If you tick only the local model, nothing is sent anywhere else, even when the local model fails.

Your own key

You can use your own OpenAI, Google AI Studio or Anthropic key instead of GeoPilot's. The request goes through GeoPilot's relay (so the key travels with it, as described above) and is served only on your key: if your key fails, this tier fails and the next ticked tier is tried, which will use GeoPilot's accounts. Untick the India region and other cloud if you want your key or nothing.

Your provider's own terms apply to these calls. In particular, a free Google AI Studio key allows Google to train on and human-review your data (see below).

2. India region (the default)

GeoPilot's hosted service, served by Google Cloud Vertex AI from the asia-south1 (Mumbai) regional endpoint, using a model that Google lists for in-region processing in India. The region can be changed by GeoPilot's operators (PIPEPILOT_INDIA_REGION); asia-south1 is the only Indian region where Google lists these models.

The request passes through GeoPilot's relay server on Vercel. This version of PipePilot asks Vercel to run that relay in Mumbai (bom1). On 12 October 2026, before that change was deployed, the relay ran in Washington DC (iad1). Until this paragraph is updated after a check on the live site, assume your question may pass through the United States on its way to Mumbai.

The India region covers where the model runs. GeoPilot's usage record of the call (your email and the first 500 characters of the question) is still stored in Sydney, Australia; see "What GeoPilot keeps".

3. Other cloud providers

GeoPilot's general fallback chain, used only when you tick it. It is served from outside India: Google's models through Vertex AI in the United States, the Gemini API or Vercel's AI Gateway (depending on GeoPilot's server configuration), and Anthropic's models when GeoPilot has them switched on. DeepSeek is part of this chain only when you also tick "Include DeepSeek" and the India region is unticked.

What each provider says

Quotes are verbatim, read on 12 October 2026.

Google Cloud Vertex AI (India region, and other cloud)

Google has renamed Vertex AI "Gemini Enterprise Agent Platform"; the old addresses redirect to the new ones.

Training: "Google won't use your data to train or fine-tune any AI/ML models without your prior permission or instruction." (zero data retention page)
Caching: "By default, Google's published Gemini models cache Customer Data (inputs, outputs, and derived data) in-memory ... has a 24-hour TTL." It "can be disabled at the project level." (same page)
Abuse monitoring: flagged prompts are "stored securely for up to 90 days in the same region or multi-region selected by the customer for their project", and "customers with a Google Cloud Master Agreement are exempt from prompt logging for this abuse monitoring by default." (abuse monitoring)
Region: "Locational endpoints : These endpoints (like us-central1 , europe-west1 ) ensure that ML processing remains entirely within the broader multi-regional or country jurisdiction associated with that region". The page lists an "India (asia-south1)" column. (data residency)
Certifications: Google lists "Generative AI on Gemini Enterprise Agent Platform (formerly Generative AI on Vertex AI)" in scope for ISO 27001, 27017 and 27018 and SOC 1, 2 and 3 (services in scope), and for ISO/IEC 42001 (ISO 42001).

Google Gemini API (AI Studio key: other cloud, or your own key)

Paid tier: "Google doesn't use your prompts ... or responses to improve our products", but logs them "for a limited period of time, solely for detecting and preventing violations of the Prohibited Use Policy", and "This data may be stored transiently or cached in any country in which Google or its agents maintain facilities." (Gemini API terms)
Free tier: Google "uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services and machine learning technologies", and "human reviewers may read, annotate, and process your API input and output." (same page)
Retention: "Google retains the following data for fifty-five (55) days for the purposes of detecting and preventing violations of the Prohibited Use Policy". (usage policies)

Anthropic (other cloud when switched on, or your own key)

Training: "By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models." (privacy.claude.com)
Retention: "For Anthropic API users, we automatically delete inputs and outputs on our backend within 30 days of receipt or generation". Inputs and outputs flagged for usage policy violations can be kept up to 2 years. (privacy.claude.com)
Region: "Inference geo: Only "us" and "global" are available." No Indian region is offered. (data residency)
Certifications: "ISO 27001:2022 (Information Security Management) ISO/IEC 42001:2023 (AI Management Systems) SOC 2 Type I & Type II". (privacy.claude.com)

DeepSeek (other cloud, opt-in only)

Location: "we directly collect, process and store your Personal Data in People's Republic of China." (DeepSeek privacy policy)
Training: the privacy policy lists, among its purposes, "To improve and develop the Services and to train and improve our technology, such as our machine learning models and algorithms." The terms of use say inputs and outputs may be used "to a minimal extent" to improve the services, with an opt-out toggle (terms of use). We could not confirm that the opt-out applies to API calls.
Retention: "We retain Personal Data for as long as necessary to provide our Services and for the other purposes set out in this Privacy Policy." No fixed period.
Law: the API terms are "governed by the laws of the People's Republic of China in the mainland." (open platform terms)
Certifications: we found none published.

OpenAI (your own key only)

Training: "data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)." (your data)
Retention: "abuse monitoring logs are generated for all API feature usage and retained for up to 30 days". (same page)
India: OpenAI offers an India data residency option for storage, not for processing, and only by arrangement with OpenAI. A standard key gets neither. (same page)
Certifications: OpenAI's trust portal lists SOC 2 Type 2 and ISO/IEC 27001, 27017, 27018, 27701 and 42001 (trust.openai.com). We could not confirm which services each covers.

Where the claims are weak

DeepSeek stores data in China under Chinese law, may use it to train its models, keeps it for no fixed period, and publishes no security certification. It is off by default and is excluded whenever the India region is ticked or only the local model is on. We recommend that government users and anyone handling government projects leave it off. An advisory from India's Ministry of Finance dated 29 January 2025, as reported by the press, told staff that AI tools "such as ChatGPT, DeepSeek etc." on office devices "pose risks for confidentiality of (government) data and documents" (The Week, 5 February 2025). We could not find the advisory on a government website.
Google's 24-hour cache and abuse logging apply to the India region unless GeoPilot's Google Cloud project has the cache switched off and the abuse-logging exemption. As of this check GeoPilot has not confirmed either; until it does, assume flagged prompts can be kept up to 90 days, in the Mumbai region.
Google's own caution: "Endpoints don't guarantee data residency or in-region ML processing." In-region processing is a commitment for the models Google lists on its data residency page, which is why the India tier uses one of those.
A free Google AI Studio key used as your own key lets Google train on and human-review your data.
Anthropic and OpenAI offer no in-India processing for a standard API key.
The relay may run in the United States until the Mumbai setting is confirmed on the live site (see India region above).
GeoPilot's usage records (your email and the first 500 characters of each question) are stored in Sydney, Australia, for every hosted tier including the India region. Only the local model tier keeps your question entirely off GeoPilot's systems.

DPDP Act 2023 notes

These notes are information, not legal advice.

What is personal data here. A network model summary is not usually personal data. Your account email is, and so may be a question or file name that names a person or a consumer. Keep personal details out of questions.
Cross-border transfer (section 16). The Act lets the Central Government restrict transfers "to such country or territory outside India as may be so notified." (Act, MeitY). We found no country notified so far, but did not search the Gazette exhaustively.
Security safeguards (section 8(5)). A Data Fiduciary must take "reasonable security safeguards to prevent personal data breach", including for processing by a Data Processor.
When it applies. Sections 3 to 17, including 8 and 16, come into force eighteen months after the notification of 13 November 2025, that is in May 2027 (G.S.R. 843(E)). The DPDP Rules 2025 were notified on 14 November 2025 with a phased start (G.S.R. 846(E)); Rule 6 lists minimum safeguards such as encryption, access control and logs, and Rule 15 allows transfers abroad subject to requirements the Central Government may set.
Government departments often have their own rules on cloud services and AI tools, separate from the DPDP Act. Check them before using a hosted tier; the local model tier keeps everything on your machine.

Engineering results

PipePilot results support professional engineering judgement and do not replace it. The engineer of record remains responsible for the design and for checking every result. No warranty is given that results are fit for any design purpose. The assistant can be wrong; check every number it gives against the model. The binding text is in the Terms of Service.

Changes and contact

GeoPilot will review this statement whenever the assistant's providers change, and at least every six months. Questions: support@smartbhujal.com.