Inside your company
If sensitive internal data is holding AI work back
Blocked work runs in the systems you already use.
Same path whether a person asks or a system calls.
Blocked workRuns in your systems
Context-Preserving Data Layer for AI
AI stops at the data it can’t touch.
AI stops at the data it can’t touch.
LLM Capsule makes it usable.
LLM Capsule substitutes the sensitive values but keeps the structure and relationships the model needs. Your approved model runs on that protected version, and LLM Capsule restores the real values through a mapping kept inside your environment.
Two starting points
First, where will you use it?
Inside your company or inside your product, LLM Capsule works the same way.
AI work that stalled on sensitive data now runs where your teams already work.
Inside your product
If sensitive customer data keeps AI features out of your product
The feature you could not ship goes into your product through the SDK or API.
Blocked featureShips inside your product
How the round trip works
Your original values stay in your environment.
Your original values stay in your environment.
Only the protected version goes to the model and back.
Substitute, run on the approved model path, reconstruct inside your environment. All three happen within one request.
Plays on its own. Click the last step to reconstruct it yourself.
NOC Assistantonline
circuit_id=CKT-77-AB12 on node edge-04 is flapping. Why, and how do I fix it?
1
Sensitive values become stand-ins
The original values stay in a reconstruction mapping in your environment.
Stand-inOriginal (stays local)
Stand-in values appear here.
2
The model runs on the stand-ins on the approved path
Only the protected version travels the approved model path.
GPT-4ClaudeLlamain-house
external · VPC · on-prem
3
LLM Capsule puts the real values back
LLM Capsule looks the values up inside your environment. No original values are sent back to the model.
Only the ⟨capsule:…⟩ form leaves your environment. Your original values and the reconstruction mapping stay inside.
Why the workflow holds
Nobody has to
Nobody has to
re-assemble the answer by hand.
What comes backanswer quality
How it runsfor operations & approval
How it compares
Three ways to handle sensitive data,
Three ways to handle sensitive data,
compared on what the work needs.
Masking removes the value, so the answer comes back incomplete. A privacy vault stores protected values, but how much task context survives varies by tool.
LLM Capsule substitutes while preserving context and restores an answer ready for real work.
Masking tool
circuit_id = ██████on node ███history ████
- Hides the value, then stops
- Structure and history are lost
Cloud privacy vault
values + mapping→ context varies by tool
- Hides sensitive values and stores them
- How much task context and relationship detail remains varies by setup
LLM Capsulecontext-preserving data layer
circuit_id = ⟨capsule:circuit_1⟩on node ⟨capsule:node_1⟩history kept
- Context-preserving substitution
- Your original values and mapping stay in your environment
- Answer returned ready for real work
Embedded DTS
When substitution alone is not enough,
When substitution alone is not enough,
DTS fills in what the task needs.
When substitution alone cannot keep the originals inside and still give the task what it needs, DTS rebuilds data with the required patterns, distributions, and relationships.
Relationships that consistent substitution already preserves stay as they are. Only the data that needs rebuilding is rebuilt.
Substitution is enoughCapsule
Customer inquiries or contract review: names, accounts and codes are swapped and the content stays intact. The same person keeps the same stand-in, and the answer comes back with real values.
Rebuilding is neededEmbedded DTS · AI-native reconstruction
Transaction pattern analysis or clinical and claims analytics: the relationships between amounts, timing and results are the answer. Swapping values one by one erases the pattern, so DTS rebuilds the data with the relationships the task needs. Rebuilt data serves analysis, training, and evaluation. Decisions on an individual customer case still run through the substitution and reconstruction path.
Not every dataset needs rebuilding.
Rebuilt data is not restored back to the originals.
Performance
Speed and quality,
Speed and quality,
measured in our own tests.
Self-tested figures. Processing time is per 2,200-character page and excludes answer generation.
0.12sper page (2,200 chars)
Values restoredthrough your own mapping
98%output similarity vs. results on original data
0.94automatic PII detection F1
Measured on
Self-tested · processing time excludes answer generation
Reconstructed through
Reconstruction mapping in your environment
Where it runs
Start with a workflow that AI can’t reach today.
Telecom & network operationsIncident response
T-Challenge 2026 runner-up, hosted by Deutsche Telekom and T-Mobile US
Works onBGP logs · circuit IDs
Returns asReconstructed incident ticket
Manufacturing & industrial operationsPlant and ICS
Works onAsset IDs · ICS alerts
Returns asUpdated field work order
Healthcare & clinical servicesClinical records
Works onClinical notes · medical record IDs
Returns asReconstructed EHR record
Public sectorCitizen services
JDC · Chungnam State University · Seodaemun-gu Urban Management Corp.
Works onCitizen requests · admin documents
Returns asDraft replies · audit trail
How it fits the platform
In production,
In production,
the data has to stay AI-ready.
Operational data changes weekly. When quality drifts, answers move between runs and last month’s approved result has nothing behind it. Syntitan, CUBIG’s AI-ready data platform, records the data version and run conditions behind each result and revalidates after changes.
-
LLM Capsule Runs blocked AI workYour blocked workflow starts running
Data you could not send to a model runs on your approved path, and the real values come back in the answer. What changes is the integration point, and its scope depends on your workflow.
-
Syntitan Records and revalidates dataEach data version is recorded and revalidated
Each data version is recorded with its run conditions and revalidated after changes, so you can reproduce the result at audit and show the basis.
Each can be adopted on its own. The operating scope grows with repeated work, changes, and revalidation.
Straight answers
What data leaves, and who signs off?
We collected the questions workflow owners and approvers ask first.
Can we use an external model like GPT-4 or Claude?
Yes. The approved model path can be an external API, a VPC endpoint or a local model. Only the protected version travels that path.
Where can LLM Capsule be deployed?
LLM Capsule can run on-prem, in a VPC or in a controlled network. You set the deployment boundary and the approved model path during architecture review.
What kinds of operational data can we start with?
Start with one representative workflow: logs, tickets, tables, PDFs, scans, or any other operational record you already work with.
Do we need to retrain the model or rewrite every prompt?
No. LLM Capsule substitutes the values before the model call and restores the result afterward, so prompts and the model itself stay as they are. What changes is the integration point, and that depends on your workflow and model endpoint.
Who decides which values are stand-ins?
Your team does, per workflow. That includes the business identifiers and references a standard PII detector would miss.
What can an approval team review?
The deployment boundary, the approved model path, the substitution policy, where the original values and the reconstruction mapping live, and how the result is reconstructed.
What should we prepare to try it on our own data?
One representative data sample, the model path you want to use, and the output your workflow needs. We run the full round trip on it.
How does this relate to Syntitan?
They are separate CUBIG products that solve different problems. LLM Capsule runs AI work that sensitive operational data had blocked, inside your environment. Syntitan records the data version and run conditions behind each result and revalidates after changes. Each can be adopted on its own, and the operating scope grows with repeated work, changes, and revalidation needs.
Integrations
Use LLM Capsule inside the tools you already run
Start with the free Slack bot. More integrations are on the way.
Communication
Slack
Ask AI in Slack, without leaving your workspace. Free to use.
Communication
Microsoft Teams
Ask Claude from Microsoft Teams, in active development and not yet available.
Productivity
Atlassian
Bring LLM Capsule into Jira and Confluence, in active development and not yet available.
More on the way
New integrations are in the works. Browse the full directory.
Explore all integrations →AI responses can be wrong or incomplete. Review anything you plan to act on.
Run one blocked workflow
inside your own environment.
See what stays in your environment and how the real values come back in the answer.