LLM CapsuleContext-Preserving Data Layer for AI
Your AI stops at the data it can't touch. LLM Capsule lets it work on that data.
Run AI on sensitive operational data and get results your workflow can use right away. The original values stay in your environment.
Available onAWS MarketplaceMicrosoft Marketplace
2-minute overviewOne workflow, start to finish
01The problem
Why enterprise AI stalls before production. The model is rarely the hard part.
The hard part is the data. Here is the same support case, handed to AI three ways.
-
01
Send it as is
StoppedGoes to AI
Case 4471 · Traffic on line 2 of OPS-WS-17 (10.4.8.17) at Riverside Plant
Policy keeps asset names, IPs, and sites from leaving. The workflow stops before the model sees anything.
-
02
Delete or mask the values
Answer can’t be usedComes back
Traffic on line 2 of is 3x normal. Ask the team to check it.
The data can go, but the answer loses who and where. Someone re-maps it by hand.
-
03
LLM Capsule
Ready to useComes back
Traffic on line 2 of OPS-WS-17 is 3x normal. Ask the Riverside Plant team to check it.
Stand-ins keep the context for the model. The original values come back in your environment.
02Capabilities
What changes when AI can work on data that can’t move raw
-
Get real results back
Original names, figures, and references are restored in the AI output, ready for reports, reviews, and client deliverables.
-
Structure and context stay intact
Values are swapped consistently, so tables, tickets, logs, and the links between them still make sense to the model.
-
Your policy decides what’s sensitive
Go beyond standard PII: device IDs, deal terms, M&A code names, OT identifiers. Change the rules without rebuilding pipelines.
- Versioned
- Scoped
- Access-controlled
- Change-logged
About 0.12 s per page
2,200-character document, self-tested, excluding answer generation.
Connects through an API
Air-gapped, on-prem, or inside your own product. One policy covers external and on-prem models.
Embedded DTS when swapping isn’t enough
Amounts, dates, and test results that lose their use when swapped are prepared for the task instead.
03How it works
Four steps, one boundary. Only stand-ins cross it.
Your systems, the original values, and the mapping stay in your environment.
Your environment
Ticketing system
Case 4471 · Unusual traffic on line 2
- Asset
- OPS-WS-17
- IP address
- 10.4.8.17
- Site
- Riverside Plant
LLM Capsule · substituted version
- OPS-WS-17→ASSET-A1
- 10.4.8.17→ADDR-A1
- Riverside Plant→SITE-A1
Mapping stays here
RestoredBack in the ticket
Traffic on line 2 of OPS-WS-17 is 3x normal. Ask the Riverside Plant team to check it.
Approved model
External or on-prem
- ChatGPT
- Claude
- Gemini
- on-prem sLLM
Model sees and answers
Traffic on line 2 of ASSET-A1 is 3x normal. Ask the SITE-A1 team to check it.
- 1 · DetectNamed-entity detection and your own rules find the values your policy marks as sensitive.
- 2 · SubstituteEach value becomes a consistent stand-in, so the record keeps its structure and meaning.
- 3 · RunOnly the substituted working version goes to the approved model, external or on-prem.
- 4 · RestoreThe answer comes back and the original values are restored through the mapping in your environment.
Policy-based routing between external and on-prem models · Retrieval for RAG with embedding and keyword search · An audit log and usage record for each run
See the full diagram04Positioning
Built to enable AI work, not to police it.
Each tool works on a different part of the same AI request. LLM Capsule changes the content itself.
-
Works onApp
Employee AI
Helps workers search, chat, and automate tasks across company apps.
-
Works onRoute
AI gateway
Manages model traffic: routing, auth, fallback, caching, rate limits, cost, observability.
-
Works onCheck
DLP
Detects, classifies, or blocks sensitive content.
-
Works onContent
LLM Capsule
Changes what your workflow sends to the model, then restores the results inside your environment.
05Proof
Customers and self-tested results
Finance · Public sector · Industrial cybersecurity
- 0.12sPer page2,200-character document, excluding answer generation
- RestoredOriginal valuesThrough the mapping in your environment
- 98%Output similarityCompared with processing the original
- 0.94PII detection F1Automated detection
Self-tested figures. Original values are restored through the mapping in your environment.
06FAQ
Frequently asked questions
What is a context-preserving data layer for AI?
A context-preserving data layer for AI lets AI work on operational data that cannot move raw. Sensitive values are consistently substituted. Document structure, relationships, and meaning stay intact, so models reason over real business context. Usable outputs are restored inside your environment, and the original values and the mapping stay there.
How do you run an LLM on operational data that cannot move raw?
LLM Capsule turns sensitive values into context-preserving substitutes before the request reaches the model. The model path can be external, but the model sees only the substituted working version. The original values and the mapping stay inside your environment, where the original values are restored in the response.
Can we run RAG or agents on documents that contain sensitive fields?
Yes. LLM Capsule substitutes the sensitive elements in your RAG sources and agent context while keeping the structure that retrieval and reasoning depend on. So RAG and agent workflows run on operational data you could not send to an external model before.
How is this different from masking or redaction?
Masking and redaction destroy the meaning a model needs, and a record full of blanks is unusable. LLM Capsule keeps the format, relationships, and document structure intact with context-preserving substitutes, so the model still understands the task and teams get answers they can act on, with real values restored internally.
How is LLM Capsule different from an AI gateway or DLP?
An AI gateway routes and manages model traffic. DLP detects and blocks sensitive content. LLM Capsule changes what crosses the model boundary: it swaps operational values for context-preserving substitutes before model execution, then restores the original values in the output inside your environment.
Where are original values restored, and what leaves our environment?
Restoration happens only inside your organization. The model path can be external, and the model sees only context-preserving substitutes. The original values and the mapping stay inside your environment. Original values are restored through your internal mapping, not through statistical recovery.
See LLM Capsule run on your own documents.
Start with one real workflow and a few of your own documents.
Available onAWS MarketplaceMicrosoft Marketplace
















