The Department of Defense Is Gearing Up for a World War Entirely in Space
Is this related to aviation?
Your decisions, running in production.
Sutro is a platform for building, optimizing, and running reliable AI Functions.
Upload an unlabeled dataset, add your feedback on the hardest cases, and Sutro engineers a prompt that reliably executes your task using off-the-shelf models.
Prompt
Optimizing…
Is this hacker news post related to aviation?
Held-out accuracy
—
F1
—
Example 1 of 5
5 expert decisions captured
Sutro can now generalize the rules based on your feedback.
What is an AI Function?
A reliable AI Function repeatedly and consistently makes the same decisions as a human expert would on a given task. Below are examples of AI Functions our customers are building.
Classifier
Score a lead based on a ideal customer profile rubric.
Judge
Give pass or fail assignments on agent traces.
Extractor
Remove personally identifiable information from medical records.
Entity resolver
Determine whether two businesses are the same.
Matcher
Determine whether a candidate is a good fit for a role.
Router
Send support tickets to the right team or person.
Relevance Judge
Determine whether a search result is relevant to a given query.
Intent Classifier
Cluster user intents within agent interactions.
Building an AI Function isn’t about scaling general intelligence. It’s about teaching a model exactly how your organization wants a task performed.
How are AI Functions run?
AI Functions can be run continuously as new inputs arrive, or in batch to process large datasets.
You can run Sutro Functions using your own model subscriptions and infrastructure, or use our batch inference service to process large datasets inexpensively.
In production today
Sutro is powering critical, production AI Functions for many happy customers today. Sutro processes millions of decisions, and billions of tokens per month, often in environments with strict data privacy and security requirements.
Applied AI teams
Agents, evals, matchers, and other production AI workflows.
Data & ML engineers
Data enrichment, entity resolution, and extraction tasks.
Training data providers
Dataset filtering, labeling, tagging, and quality assurance.
Request customer case studies →“Sutro saves our team countless hours, and it gives us the invaluable ability to measure, optimize, and prevent regressions against our domain expertise. It’s a must-have for serious AI developers.”
Simple pricing
Platform access starts at
$500 / month
(includes $100/mo inference credits)Enterprise and self-hosted plans are tailored to your needs and scale.
See full pricing →Batch inference
Machine-time pricing means you only pay for the compute your workload actually consumes.
Estimated price
Number of records
1 million
Frequently asked
Is this an evals product?
Yes and no. Sutro helps evaluate quality and reliability, but unlike other evals products, it directly influences model/agent behavior from feedback.
Can Sutro use image inputs and other modalities? Can it call tools?
Yes. Sutro supports images, PDFs, and web search capabilities today. Additional modalities and custom tool calling are available on a per-request basis.
What models does Sutro Functions use?
Functions are typically model-agnostic, and we support a mix of open-source and proprietary models. It's common that small-open source models outperform larger proprietary models when building AI Functions. Sutro helps automatically find the best model for your needs.
Can I self-host Sutro? Do you support BYOC and BYOK?
Yes. Sutro can be managed as SaaS or entirely self-hosted. You can bring your own provider keys and cloud credentials in either case.
How does Sutro compare to prompt engineering, fine-tuning, or RL?
AI Functions accumulate edge cases by nature of the volume they run at and the diversity of possible inputs. Sutro systematically discovers these edge cases and learns generalized rules optimized to cover as many of them as possible. Manual prompt engineering is typically not scalable in these scenarios.
Sutro Functions are designed to be extremely data efficient and portable. Functions learn in-context, meaning no weight updates are needed and can easily be updated. Fine-tuning and reinforement learning typically require more data and require retraining for updates, making them more rigid and less portable.