Prompt Interlok

Prepare customer inquiries for AI-assisted review.

Customer inquiries, whether emails, quotation requests, or project notes, routinely carry client names, project codes, and commercial detail. Prompt Interlok replaces detected sensitive details, preserves useful context and adds model-assisted checks. The result needs review before further use; replacement does not guarantee anonymity.

Try the demo Read the project article

How it works

From raw inquiry to prepared text and review findings.

Rules and pattern matching identify details to replace or preserve. Detected sensitive spans become descriptive placeholders. A model stage checks the prepared text for remaining sensitive information and classifies the inquiry.

Both rules and model checks can miss information. Use synthetic examples in this demo. Processing services may receive original details in review records, even when those details are removed from model-facing text. The project article explains these data paths and the limits of the demo.

Rules first

Rules and pattern matching identify sensitive details such as client names, project codes, monetary figures and competitor mentions for replacement. This stage can run without a model, with optional model-based recognition to help identify names and organisations.

Typed placeholders, with useful context

A person’s name becomes a placeholder such as [PERSON_CLIENT], indicating what was removed without showing the name. Useful context, such as the role “Technical Buyer” or a technical protocol, is preserved according to the configured rules.

Model check and classification

A model receives prepared text and selected signals. It looks for potentially sensitive details and classifies the inquiry by type, urgency and scope. The findings should be reviewed by the user before further use.

Review the changes

Each replacement is recorded with the rule that triggered it and its position in the source text, so you can check what changed and why.

Prompt Interlok

Sanitize, verify, review.

Paste an inquiry, or load one of the examples. The tool runs a deterministic sanitization pass that removes identifying information by rule, then a small LLM checks what remains and adds a triage read. After both passes, the sanitized version is something you can send to a frontier cloud LLM that you would never trust with the raw inquiry.

Sanitize

Start here

Deterministic pass

Replace sensitive spans with typed placeholders.

Select an example or paste an inquiry, then run the sanitization pass.

Verify

Locked

Small LLM check and triage

A small LLM reads only the sanitized output.

Sanitize the inquiry first.

Review

Locked

Read and copy the output

Inspect the sanitized output. Copy it when you are ready.

Verify the sanitized output first.

Stage 1, Sanitize

Choose an example or paste a sample inquiry, then run sanitize.

Example

Load an example, or paste a short fictional inquiry. This public demo uses a fast low-cost API model for verification and is limited to 1500 characters. Do not paste real client data here.

Public demo mode815 / 1500 characters

What the sanitization pass did

Run sanitize to see what was replaced, flagged, and preserved.

Before and after

Original

Subject: Inquiry for pull-wire safety switches on new waste-processing conveyors

Hello,

We are commissioning three new conveyors for a waste-processing plant near Łódź and need your recommendation for pull-wire emergency stop switches.
The longest conveyor is 42 m between end pulleys. The installation area is wet, dusty, and exposed to washdown, so we need a corrosion-resistant version with minimum IP65.
We initially looked at ACM-PC-LR, but please advise if another model from your range would be better together with the required mounting accessories.
Please also confirm current lead time for delivery in Q3.
Our internal project reference is WP-2/A6; please include it in your reply.

Best regards,
Marta Kaczmarek
Project Engineer
Mazovia Resource Recovery Sp. z o.o.
m.kaczmarek@mazovia-recovery.example

Sanitized

The sanitized version will appear here after the pass runs.

How this demo is configured

What the rules and models actually do

The interaction above runs against a fixed demo configuration: a deterministic rule set, a small verification model, an approved whitelist, and a documented trust boundary. This section explains what stays local, what reaches the verification model, and which live runtime details are shown for the current run.

Where the verification model runs

Privacy depends on where the verification model runs. The strictest setup keeps the sanitized check on local hardware capable of running a small LLM under the operator's control. A hosted setup can also work when it runs under an account the operator controls and policy allows sanitized text to leave the browser. This public demo uses a hosted small model so the workflow stays responsive on an open website. In every setup, only sanitized text reaches the verification model; the raw inquiry does not.

What's running right now

Run the verification pass to see live runtime details for this session.