We’ve partnered with NEAR AI to bring Sage, our decision model, to its confidential computing infrastructure.
Sage now runs inside a trusted execution environment (TEE). Teams can use it to classify sensitive records, evaluate confidential information, and make structured decisions that drive application logic while protecting the data during inference.
For enterprises processing customer conversations, internal documents, operational records, or proprietary research, this brings confidential computing to the decisions embedded in their data pipelines.
Enterprise automation requires access to sensitive context
Many business processes depend on judgments that conventional rules cannot easily express.
Does this document contain information relevant to an investigation? Which category describes this customer request? Does this record meet the criteria for further processing? Should this case proceed automatically or require human review?
Answering these questions requires understanding the information in each record. The inputs vary in wording, format, and completeness. Rules based on exact matches become difficult to maintain as that variation grows.
Sage lets developers express these judgments as questions with structured answer formats. The application supplies the relevant context, receives a decision and confidence signal, and uses the result to determine what happens next.
At scale, this becomes part of the data pipeline: classify incoming records, filter irrelevant material, prioritize work, and escalate cases that require attention.
The challenge is that the context needed to make those decisions often includes information a company cannot expose to an infrastructure operator.
What Sage does inside a data pipeline
Sage, the Levanto decision model, is developed by Levanto Labs for machine-to-machine decision making. It evaluates context and returns structured outputs with calibrated confidence signals.
Developers can use Sage to assign categories, evaluate conditions, rank items, and select among defined outcomes. Those outputs become inputs to ordinary application code. The Choice decision guide explains how to select one option. The Tags decision guide explains how to evaluate each label.
Consider a pipeline processing confidential customer correspondence. For each message, an application could ask Sage to:
- Classify the request against the company’s operational categories.
- Identify whether it meets defined escalation criteria.
- Evaluate whether the available information is sufficient for the next step.
- Determine whether a specialist needs to review it.
The application can then update a record, create a task, continue processing, or request human review.
Confidence thresholds give developers another control over automation. Teams can evaluate those thresholds against representative data and decide which cases qualify for automatic processing.
This makes semantic judgment a reusable operation within a pipeline, with explicit inputs, outputs, and downstream behavior.
What NEAR AI adds: protection during processing
Running Sage on NEAR AI brings confidential execution to these classification and decision workloads.
A trusted execution environment is a hardware-isolated environment designed to protect code and data during processing. Encryption in transit protects information as it moves between systems. Confidential computing addresses the period when the model is using that information.
NEAR AI’s confidential infrastructure uses Intel TDX and NVIDIA confidential computing. Its architecture includes cryptographic attestation: evidence that can be checked to verify the execution environment. NEAR AI also supports independent verification through Intel Trust Authority.
This gives engineering and security teams a verifiable execution boundary for sensitive inference workloads.
Sage provides the classification or decision. NEAR AI provides the infrastructure that protects its execution.
Large-scale classification of confidential data
Classification is a recurring operation in enterprise systems. Organizations need to organize records, identify relevant information, prioritize exceptions, and determine which process should handle each item.
Sage can evaluate the meaning of a record against developer-defined categories. This supports inputs whose wording varies across customers, teams, and source systems.
Running that classification inside a TEE makes the approach applicable to workloads containing sensitive business information, including:
- Customer operations: Categorize support conversations and identify cases that meet escalation criteria.
- Internal document processing: Classify reports, correspondence, and agreements for the appropriate workflow or reviewer.
- Proprietary research: Assess incoming material for relevance before committing additional processing resources.
- Operational review: Flag records that require investigation under a team’s defined criteria.
The value compounds when the same judgment must be applied across a large dataset or a continuous stream of incoming records. Developers define the decision once, evaluate its performance, and integrate it into the processing workflow.
Filtering data before expensive downstream processing
A classification decision can also determine whether additional work is necessary.
Many pipelines send every incoming item through research, enrichment, analysis, or generation. A decision earlier in the pipeline can identify which records justify those operations.
For example, Sage can assess whether an incoming document contains relevant information before the application schedules a more expensive analysis. It can distinguish routine cases from exceptions before allocating specialist review.
The resulting savings depend on the workload: how much material can be filtered, the cost of the decision step, and the downstream work avoided.
For confidential datasets, that initial evaluation needs the same protection as the analysis that follows. Sage on NEAR AI allows teams to perform it inside a protected execution environment.
Automating decisions on sensitive information
Classification can feed directly into controlled automation.
A document-processing system might continue only when a record meets defined criteria. A customer operations workflow might require approval for particular requests. An internal review system might prioritize cases according to their content and the organization’s policies.
A typical implementation follows four steps:
- Define the decision. Specify the question, permitted outcomes, and context Sage needs.
- Evaluate the record privately. Run Sage on NEAR AI’s confidential infrastructure.
- Apply application controls. Check the result against validation rules and confidence thresholds established through evaluation.
- Execute the next step. Continue processing, update the record, or escalate for review.
Application code retains control over execution. Teams can record the decision, model version, policy version, and resulting action in their own controlled audit systems.
Attestation and application records serve different purposes. Attestation provides evidence about the execution environment. Application records document what the workflow decided and did.
Protecting the complete data workflow
Confidential inference protects the workload within the TEE. A production system also needs appropriate controls around ingestion, storage, logging, and downstream processing.
For example, a pipeline should avoid copying sensitive inputs into unrestricted application logs. Any service receiving a record after classification must be authorized to handle that data.
Running Sage inside a TEE gives teams a protected decision step within that broader architecture. It supports a security design that accounts for how information moves through the entire workflow.
Build confidential data workflows with Sage
Sage is now live on NEAR AI’s confidential infrastructure.
The partnership enables teams to bring structured AI decisions into sensitive enterprise workflows, from large-scale data classification to filtering pipelines and automating operational decisions.
Contact Levanto Labs about a confidential Sage deployment, or read the Sage decision API quickstart to explore its decision interface.
