Architecture

Runs inside your environment, without touching the systems you already have

LLM Capsule sends approved AI a substituted version of your data, then reconstructs the real answer inside your network. This page has what a security, data, or platform team needs to judge it: the zones and data path first, then components, deployment, and integration.

View Pricing Run a sample proof
Where it sits

LLM Capsule runs inside your network, beside your existing systems

Documents are not sent out to a third-party service. LLM Capsule is deployed inside your network, and your existing systems call it there.

internal

Your internal systems

No replacement, no migration. LLM Capsule reads the data where it already lives.

  • ERP
  • CRM
  • ticketing
  • document management

LLM Capsule

Sensitive items leave as substituted values, and the returned output is reconstructed to its original values here.

  1. Detect
  2. Substitute
  3. Run
  4. Reconstruct

Original values and the reconstruction mapping never cross this line.

boundary
  • outbound substituted values
  • inbound model output
external

Approved AI models

Only models your organization approved. What arrives here is the substituted document.

  • ChatGPT
  • Claude
  • Gemini
  • Perplexity
  • any LLM API
  • on-prem sLLM

So the record left outside your network contains no original values.

Zone overview

Four zones, and the boundary between them

The path runs in order. Operational systems and original values stay in your environment. Only a substituted, structure-preserving copy (the working version) crosses the boundary, and the result is reconstructed where the work happens.

Zone 1

Operational systems

Records and original values stay in the systems that already run the work. LLM Capsule connects to ERP, CRM, tickets, documents, databases and RAG. No replacement, no data migration.

  • ERP · CRM · ticketing
  • Documents · databases · RAG
  • REST · gRPC · JDBC · Graph API
Zone 2

Structure-preserving substitution

Sensitive markers become a protected working version, while tables, fields and references stay usable. Only that version moves forward. Original values and the reconstruction mapping stay here.

  • Sensitive markers identified
  • Tables · fields · references retained
  • Values · reconstruction mapping stay inside
Zone 3

Policy-based routing

Your policy routes the working version by workflow, data type, permissions and model approval. The work runs on an approved external or on-premise model. Original values never enter this zone.

  • Workflow · data policy applied
  • Approved external model
  • On-premise model
Zone 4

Reconstruct into the workflow

LLM Capsule combines the model response with the local mapping and reconstructs it in its original context. The result returns straight to tickets, reports, records or workflows, with each reconstruction logged for traceability.

  • Reconstructed through local mapping
  • Tickets · reports · work records
  • Reconstruction recorded
Data boundary Working version crosses Original values and reconstruction mapping stay inside

Technical review · implementation path

Review the implementation path behind one operational request

Source interfaces, the working version, the approved route, and local reconstruction, all shown as one reviewable implementation surface.

Operational systems

Reads directly from the operational systems that are already in place.

CONNECTED SYSTEMS

ERPSAP / Oracle CRMSalesforce RAGVector DB LegacyOracle TicketsJira / SN DMSSharePoint

RAW INPUT FIELDS

Customer
David Lawson
ID
Ticket
CS-4203
CD
Free text
David Lawson, 010-4102-8823, shipment delayed 3 days, requests refund
BD

ID identifier · CD reference code · BD unstructured body

Integration surface · REST · gRPC · JDBC · Graph API

No system modification Connects to existing ERP, CRM and legacy systems with a single API-call addition. Nothing is refactored or migrated.

Structure-preserving substitution

Detection and substitution preserve the task structure.

SUBSTITUTION FLOW

INPROCESSOUT

Customer(capsule:person_1)ID

Ticket(capsule:champ_1)CD

Free text[context-preserving]BD

capsule:id is the substituted-value format that stands in for the original inside the capsule.

DP ENGINE

  • epsilon-DPActive
  • Laplace noise
  • k-anonymity enforcement
  • Semantic substitution
  • Free-text NER detection
Structure-preserving Detection and substitution keep tables, fields and references usable. Only the working version crosses the boundary; original values and the reconstruction mapping stay in this zone.
DATA BOUNDARY Working version crosses Original values and reconstruction mapping remain inside

Policy-based routing

The organization selects the approved execution route.

LLM PIPELINE

PATH A · EXTERNALPublic cloud · region-hosted
PATH B · ON-PREMPrivate · on-prem

The approved route can change without altering the working-version contract.

PRIVACY ISOLATION MODULE

  • Privacy descriptorPDESC
  • Isolated computation zoneISOLA
  • Output validation

EU GDPREU Region

No direct access to source DB The execution route never touches the source database. It receives only the structure-preserving working version.

Reconstruct into the workflow

The response is reconstructed in the environment where the work began.

RESPONSE TO ORIGINAL CONTEXT

Customer
David Lawson
ID
Ticket
CS-4203
CD
Free text
David Lawson, 010-4102-8823, shipment delayed 3 days, requests refund
BD

RECONSTRUCTION

  • Local mapping lookup
  • Original value restore
  • Context re-binding
  • Output validation

OUTPUT PROPERTIES

  • Original values restored
  • Zero external exposure
  • Context preserved
  • Business-ready output

Auto restoreLocal onlyData residency

Local reconstruction The reconstruction mapping exists only inside the customer environment; the result returns to the originating tool immediately.
Business-ready result returns to the existing operational workflow
Five implementation components

Five technical components behind four workflow functions

These components implement the four functions on the product page. Detection and substitution look like one step from outside. Here they split into two, so you can configure, observe and audit each on its own.

Structure-preserving substitution · one step from the outside

  • Component 01

    Detection Engine

    Finds sensitive data in structured fields or free text.

    PII + customer-defined markersBeyond regex: semantic + context

    98.1%detection accuracy

  • Component 02

    Substitution Engine

    Swaps sensitive values only. The shape around them stays intact.

    ε-DP · Laplace noise · k-anonymity · NERTables · hierarchies · references intact

What makes it operable at scale

  • Component 03

    Policy Definition

    Sensitivity rules keep changing. Policy keeps pace without breaking past work.

    Versioned · scoped · RBAC'dTime-shifting: archive yesterday, enforce todayPer-team, per-workflow scope
  • Component 04

    Reconstruction Engine

    Brings the real answer back to where the work started.

    Local reconstruction-mapping lookupContext re-bindingValidated, ready in the originating tool
  • Component 05

    Operational Records

    Every step from detection to reconstruction leaves an auditable record.

    Policy version · model · latency · outcomeGDPR · HIPAA · SOX
Implementation principles

Six engineering principles behind the workflow

These are technical constraints and implementation criteria, not a second set of product features. Together, they make the customer-facing capabilities reliable in production.

  • Pillar 01
    Reads free-text, not just structured PII Free-text fields like a CS ticket mix customer names, contact information, and claim narrative with no fixed structure, so simple PII guardrails cannot safely process them.
    • Free-text + structured fields together
    • Beyond pattern-matching: semantic + context
    Zone 2 · Substitution
  • Pillar 02
    No modification of existing systems Existing enterprise systems are never refactored. Connection is a single API-call addition, and operations teams keep using the tools they already have.
    • REST / gRPC
    • Substitution and reconstruction handled by the adjacent LLM Capsule layer
    Zone 1 · Source
  • Pillar 03
    Beyond simple substitution: differential privacy The Substitution Layer does more than swap values in place. It applies differential privacy to lower the re-identification risk of the working version itself.
    • epsilon-DP
    • Laplace noise
    • k-anonymity
    • NER detection
    • Structure preserved beyond a plain swap
    Zone 2 · Substitution
  • Pillar 04
    External AI only sees the substitute External AI sees only the substituted version, so the original operational data never crosses that boundary.
    • Structure-preserving substitution + DP combined
    • A leak on the external side cannot be reconstructed
    Boundary · Zone 3
  • Pillar 05
    In-environment auto-reconstruction LLM Capsule reconstructs the substituted values in the AI response back to their originals inside the organization only, never outside it.
    • Reversing a substitute needs the in-house mapping
    Zone 4 · Reconstruction
  • Pillar 06
    Governance, policy & domain context retained Organizational policy, permissions, and domain context drive the entire AI processing path, and governance stays inside the organization end-to-end.
    • What gets substituted, where it routes, how reconstruction is audited
    • Path A · external
    • Path B · on-prem
    Zone 3 · Routing

Each principle addresses a concrete architectural failure mode: loss of structure, system rewrites, context loss, an uncontrolled external path, manual reconstruction, or policy drift.

Two execution paths

One architecture. Two paths. Policy-driven per workflow.

The same LLM Capsule instance can route different workflows through different paths, all under one governance framework.

PATH A · External

Approved external LLM, capsule data only

For workflows where the rules allow a substituted, differentially-private version to reach an external model under the usual contracts (DPA, SCCs).

  • Routes to ChatGPT, Claude, Gemini, Perplexity, or any LLM API
  • Only the capsule travels. The original data never does.
  • In-region endpoints supported (EU-hosted for sovereign AI)
  • Best for: NOC RCA, claims classification, summarization
PATH B · On-prem

On-prem local lightweight model

For workflows where no external endpoint is acceptable: classified data, operational-technology (OT) systems, or tightly regulated healthcare records.

  • Quantized model on internal GPU (vLLM-served)
  • Zero external transmission, with a fully air-gapped option
  • Same LLM Capsule instance, same audit, same policy framework
  • Best for: defense, classified workflows, strict sovereign AI
In-environment integration

Reads the systems you already run without modifying them

These are the operational systems LLM Capsule reads from directly. Each connects with a single API call, and none of them change on their side.

ERP System
SAP / Oracle
CRM
Salesforce
Ticketing
Jira / ServiceNow
DMS / ECM
SharePoint
Legacy DB
Oracle / MSSQL
RAG Pipeline
Vector DB
Agent
LangChain / AutoGen
MCP Server
Claude Desktop / Cursor
  • Zero changes to existing systems
  • Connects with a single API call
  • Original values stay inside the environment
  • Runs next to your systems, on-prem or in-VPC

Invocation interfaces, matched to each stack

REST / gRPC
Modern ops tools, RAG pipelines, custom orchestrators.
JDBC / ODBC
Legacy DBs (Oracle, MSSQL, DB2), via stored procedure or job step.
Graph API
DMS/ECM (e.g. SharePoint), triggered by document events.
On-prem API
Same contract across air-gapped, hybrid, or VPC.
Embedded SDK integration
Library-level integration for ISVs and platform vendors.
Slack App integration
For teams using Slack as the operations UI.
Deployment modes

Six deployment modes, matched to your environment

LLM Capsule runs inside the customer environment in every mode. Path A and Path B execution choices apply across all six.

Air-gapped on-prem

Fully internal. No external network. Path B only. Defense, classified, OT.

On-prem hybrid

Internal LLM Capsule + approved external LLM. Path A for most workflows, Path B for sensitive subset.

VPC / private cloud

Customer's cloud VPC. LLM Capsule and reconstruction mapping stay in the tenant; the LLM call goes to an in-region endpoint.

AWS / Microsoft Marketplace

Listed and procurable through AWS Marketplace and Microsoft Marketplace. VPC deployment, cloud billing integration.

Embedded SDK deployment

For ISVs and platform vendors building LLM Capsule into their own product. Library-level integration that ships inside the host application.

Slack App deployment

For teams using Slack as the operations UI. LLM Capsule runtime stays in the customer environment; the Slack App is the surface that invokes it.

Telecom-grade deployment

A separate operator-grade topology (NFV, container, multi-region) that composes with the modes above rather than replacing them.

  • Validated at SK Telecom
  • Recognized at Deutsche Telekom T Challenge 2026
Audit & approval

The evidence security and legal sign off on

Once the working team is convinced, security and legal sign off. LLM Capsule records audit logs, policy history and exposure control as you run. Full documentation lives in the Trust Center.

Audit log

LLM Capsule logs detection, substitution and reconstruction as events, with retention set by policy.

Policy history

Every change to detection or routing policy keeps the prior version and when it changed.

Zero Exposure

Original values and the reconstruction mapping never leave the customer environment.

Certifications, compliance framework mapping, and DPA documentation are covered in the Trust Center. Visit Trust Center

See this architecture run in your own environment.

Bring your deployment constraints, regulatory profile, and one real workflow. We’ll show the data layer running in your environment in 30 minutes.

Visit Trust Center View pricing Run a sample proof