LEVEL 01
Detection & Tokenization
Catches names, phones, emails, passport IDs, tax numbers, national IDs, account numbers. Swaps them for placeholders, keeps the mapping locally.
Covers 80% of scenarios in LOOP, ECHO, ATLAS, MOONLIGHT.
Your data doesn't leave your stack raw. A local gateway masks personal data before it reaches the LLM and decrypts the answer back on your side.
Not a standalone product. Part of every DNSHN module.
When LOOP analyzes deals — customer names don't fly to the cloud. When ECHO replies — the model sees "customer #4713", not Maria Sokolova. When ATLAS enriches ICP — the names of your target accounts don't settle in someone else's context window.
It happens without your involvement: GATE sits between DNSHN products and LLM providers (Claude, OpenAI). It catches personal data in text and tables, replaces it with placeholders, and after the model responds — puts everything back on your side.
CRM, Sheets, ad accounts — as you have them. We don't migrate anything.
Names, phones, emails, IDs, tax numbers, amounts — into placeholders. Mapping key stays local, with you.
Claude/OpenAI sees the task shape, not the content. Their training set doesn't see it either.
GATE reads the response on your side, puts names and numbers back in, hands it to the DNSHN product.
Customer: Maria Sokolova Phone: +1 415 555-11-22 Deal amount: $12,000 Note: pushed payment to next Monday.
Customer: [PERSON_1] Phone: [PHONE_1] Deal amount: $[MONEY_1] Note: pushed payment to next Monday.
Model responds: "[PERSON_1] — cancel probability low, 3-day push within norm. Ping a day before." GATE puts it back: "Maria Sokolova — cancel probability low…"
Catches names, phones, emails, passport IDs, tax numbers, national IDs, account numbers. Swaps them for placeholders, keeps the mapping locally.
Covers 80% of scenarios in LOOP, ECHO, ATLAS, MOONLIGHT.
Amounts, CPAs, budgets get replaced with numbers of the same order of magnitude. The model can reason about scale; the exact value only unlocks on your side.
Lets the LLM tell "large deal" from "small", without seeing the actual figure.
Shipping a raw database to the cloud while keeping analytic validity. Differential privacy, secure MPC, homomorphic encryption.
If that's really your job — you need an on-prem LLM, not a gateway. We'll say so upfront.
Tabular data with clear structure: deals, leads, payments, ad accounts. The LLM answers a specific question — the answer is meaningful.
Client conversations: names and phones get masked, content stays — ECHO responds in brand voice without knowing who exactly it's talking to.
Free-form reasoning about the person ("who's this customer, how to close them"). Without a name the model loses gender, role, emotional cues. Answers turn averaged.
Exact arithmetic on masked amounts: "sum this column" gives an order of magnitude, not an exact figure. For financial audit GATE won't replace on-prem.
We don't sign compliance certificates. We substantially cut the leak surface — but full compliance needs an on-prem LLM plus legal wrapping.
A technical layer + documentation: what gets masked, what doesn't, where the mapping lives, how to revoke keys. Enough to pass a security review.
Deploy — docker-compose, one command. Data doesn't leave your perimeter except in masked form. Mapping keys stay on your side; we don't see them.
Within 24 hours we'll say what GATE will mask in your data — and what it won't — with an example on a real fragment (no PII in the answer).