AgentAnywhere Swaraj
Aavaran9 min readSee Aavaran →

Mask it before the model sees it.

Every AI agent you deploy is a new reader of your customers’ data. Aavaran makes sure that reader only ever sees the masked form — and leaves a signed receipt for every value it hid.

AgentAnywhere Research

· Updated

Animated diagram: a support message containing an account number, card number and phone number travels to a perimeter line, where Aavaran replaces each value with a mask and logs a signed Trust Receipt; only the masked message continues to the model.
FIG.28Only the masked form crosses the line — the raw account, card and phone values are masked before the model, and a signed Trust Receipt records the masking.

Watch it happen: only the masked form crosses the line.

This is not a diagram. It is Aavaran’s own detection and masking engine running on a support message. Sensitive values are found inline and replaced with masks that keep only a small recognisable part — the last four digits of a card or phone number, the first letter of an email’s name and of its domain, the first two and the last character of a PAN. Other fields are masked in full. Only the masked form crosses your perimeter to the model. It is designed so that the raw value does not leave the box.

Live — running in your browser

INBOUND — a support message with sensitive values

Hi, I'm Anita Rao. My email is a***@e***.comemail and my card ************1111card was charged twice. My PAN is AB***FPAN.

your perimeter
crosses to the model:a***@e***.com************1111AB***F
● sealed3 of 3 sensitive values masked before egressreal Aavaran engine output · raw values never left the box

It is 11 pm in the contact centre

A customer has typed her account number, her card number and her mobile into the chat, because that is what customers do when they are worried about a double debit. The agent handling her — increasingly, an AI agent — needs to understand the problem. It does not need her account number to do that. It needs to know that there is an account, and which transaction she means.

For years the honest answer to “does the model see the raw data?” was yes, with a policy document saying it shouldn’t. The model sits at the end of the pipe; whatever went into the pipe reached it. Logs kept copies. Embeddings kept fingerprints. A vendor’s support engineer could, in principle, read it.

Aavaran moves the decision to the one place it can actually be enforced: the perimeter, before the call goes out. The customer’s message reaches the model with the account number masked in full and the card shown as ************8813, its last four digits. The model does its job. The raw values stayed home.

Classify. Mask. Enforce. Prove.

Four verbs, in the order they happen on each request.

  • Classify

    Detect PII, financial and regulated fields inline, on each request it handles — contacts, account and card numbers, health identifiers, and the fields your own policies designate.

  • Mask

    Reversible tokenization, switched on per deployment, gives a deterministic token your downstream systems can keep matching on, with a format-preserving scheme available as an option. Irreversible redaction replaces the value so it cannot be recovered: in the full style nothing of it remains, and the default partial style keeps only a small part.

  • Enforce

    Designed to sit in front of the model, as an OpenAI-compatible proxy or as a mask-before-inference call that a gateway or agent runtime makes. Not something a prompt can talk its way around.

  • Prove

    Every mask and unmask becomes a Trust Receipt — signed, versioned, exportable as JSON or PDF.

Reversible when you need it back. Irreversible when you don’t.

The right kind of masking depends on who is downstream. Aavaran lets policy choose per field.

Tokenize — operations that must complete

A bank’s KYC operations agent has to look up the account after the model has understood the request. Reversible tokenization keeps a deterministic token in the model’s view and lets an authorised system — never the model — resolve it back inside your perimeter.

The unmask is itself a logged, policy-bound event. Nobody resolves a token without leaving a receipt.

Redact — data that should not exist downstream

A defence programme triaging incident reports, or a space agency sharing operational logs with an external model for summarisation: names of personnel, locations and identifiers should not survive the trip. Irreversible redaction removes them before egress, and there is no path back.

The receipt records who, what, under which policy and when, so there is a record that the redaction happened and under which rule.

Where it is doing the work

These are the deployments we are asked about most. None of them required replacing the agent or the model already in place.

  • Contact centres and BPOs

    Chat and voice transcripts carry everything a customer chooses to type or say. Aavaran masks them before the assist model, the summariser and the QA scorer see them — and gives the client a receipt trail per interaction.

  • PSU and private banks

    KYC operations, complaint handling and collections all want an AI assist and none of them want account or card data in a model log. Tokenize for the operations that must complete; redact for the rest.

  • Insurers and NBFCs

    Claims documents mix health identifiers with financial ones. Field-level policy lets a claims agent reason over a masked document while the identifiers stay inside your boundary.

  • Defence and strategic programmes

    Irreversible redaction of personnel, unit and location identifiers before any model — including one running air-gapped inside the wire — sees the document.

  • Space and mission operations

    Operational logs and anomaly reports summarised by a model without the contractor, site and personnel details ever egressing.

  • Anyone with an auditor

    Trust Receipts turn “we mask sensitive data” from a policy statement into an exportable record — who, what, which policy, when.

What Aavaran is not

Aavaran is not a prompt-injection shield — that is Kavach. Its masking policy is created in Aavaran itself. And it is not a promise that every possible sensitive field in every language will be caught on day one: classification is configured and measured with you, and the fields your policy designates are the fields Aavaran enforces.

It is also not a certification of your deployment. SOC 2 and ISO 27001 are held certifications. HIPAA and GDPR readiness is independently assessed. Aavaran is designed to support India's DPDP Act and the RBI FREE-AI framework. These describe the AgentAnywhere platform and how ShepHertz operates it; they are not a certification of your own deployment.

Start where the data is

Aavaran stands on its own in front of whatever you already run — your agents, your models, your vendor’s models — with its own masking rules and its own receipt trail. No platform migration. When you are ready, it is designed to sit in front of the platform's models too, as an OpenAI-compatible proxy or as a mask-before-inference call that a gateway or agent runtime makes; Swaraj is designed to run it air-gapped, and every receipt lands in the same audit record as the rest of the platform.

The only real prerequisite is a decision about which fields matter to you. We will help you make it, and then we will show you the receipts.

Frequently asked questions

How do you stop an AI agent from seeing customer PII?

Mask it before the call leaves your perimeter. AgentAnywhere Aavaran classifies PII, financial and regulated fields inline on each request it handles and replaces them with masks — reversible tokens or irreversible redactions — so the model, tool or embedding service only ever receives the masked form. Aavaran is designed to sit in front of the model, as an OpenAI-compatible proxy or as a mask-before-inference call that a gateway or agent runtime makes, so that raw data does not reach a model unmasked, with every mask and unmask logged.

What is the difference between tokenization and redaction in Aavaran?

Reversible tokenization, which is switched on per deployment, replaces a value with a deterministic token that downstream systems can keep using to match records and an authorised system can resolve back inside your perimeter, with a format-preserving token scheme available as an option; every unmask is logged. Irreversible redaction replaces the value so that it cannot be recovered: in the full style nothing of it remains, and the default partial style keeps only a small part, such as the last four digits of a card. Policy chooses per field.

What is a Trust Receipt?

A Trust Receipt is Aavaran's audit artifact. Every mask and unmask is logged — who, what, under which policy, and when — then signed and versioned. Trust Receipts can be exported as JSON and PDF for auditors.

Can Aavaran be used with the AI models and agents we already have?

Yes. Aavaran deploys standalone in front of your existing agents and models, including third-party ones, with its own masking rules and receipt trail. It is also designed to work with the AgentAnywhere platform, in front of the platform's models, as an OpenAI-compatible proxy or as a mask-before-inference call that a gateway or agent runtime makes; Swaraj is designed to run it air-gapped.

Which compliance frameworks does Aavaran align with?

SOC 2 and ISO 27001 are held certifications. HIPAA and GDPR readiness is independently assessed. Aavaran is designed to support India's DPDP Act and the RBI FREE-AI framework. These describe the AgentAnywhere platform and how ShepHertz operates it; they are not a certification of your own deployment.

TopicsPII masking for AI agentsdata masking before LLMde-identification AI contact centreDPDP Act AI data protectionreversible tokenization AI

Written by

AgentAnywhere Research

The team that builds the platform and the models

AgentAnywhere Research writes about the platform, the model families and the trust layer we build and run in India. Where a figure is ours, it says what it covers; where something is a demonstration, it says so.

All articles →