# Sutro

> Sutro is a platform for building and running consistent AI Functions by optimizing prompts from human feedback on unlabeled datasets.

Sutro is a platform built by Skysight, Inc. that helps applied AI teams create reliable, repeatable AI Functions — structured AI tasks that consistently replicate expert human judgment at scale. It is currently in production use, serving customers in environments with strict data privacy and security requirements, according to the Sutro website.

## What It Is

Sutro occupies the space between raw prompt engineering and full fine-tuning. An "AI Function" in Sutro's terminology is a discrete, repeatable AI task — a classifier, a judge, a PII extractor, an entity resolver, a router, or an image classifier — that reliably produces the same output a human expert would. Users upload an unlabeled dataset, provide feedback on the hardest edge cases, and Sutro engineers an optimized prompt that generalizes those decisions across off-the-shelf models. The platform is model-agnostic, supporting a mix of open-source and proprietary models, and automatically selects the best model for each function.

## How AI Functions Are Built and Run

The core workflow has three steps: upload data, annotate difficult cases, and let Sutro optimize the prompt. Once a function is ready, it can be invoked in two modes:

- **Single inference** via `POST /v1/run/{function-name}` for real-time, event-driven use cases
- **Batch inference** via `POST /v1/run-batch/{function-name}` for cost-efficient transformation of large datasets

Functions improve over time — users can return to learn from new data or re-optimize against newly released models. The platform also surfaces low-confidence results to an annotation queue, enabling continuous refinement.

## Supported Use Cases

The Sutro site highlights five primary AI Function archetypes:

- **Classifiers** — e.g., scoring leads against an ICP rubric or detecting vehicles in bike lanes from images
- **Judges** — evaluating agent traces for pass/fail quality
- **PII/PHI extractors** — redacting personally identifiable information from medical records
- **Entity resolvers** — determining whether two business records refer to the same entity
- **Routers** — directing support tickets or inputs to the correct team or workflow

## Target Audience

Sutro is positioned for three overlapping audiences: applied AI teams building agents, evals, and production AI workflows; data and ML engineers handling enrichment, extraction, and entity resolution; and AI training data teams doing dataset filtering, labeling, tagging, and quality assurance.

## Deployment and Infrastructure

Sutro can be run as a managed SaaS or entirely self-hosted. It supports bring-your-own-cloud (BYOC) and bring-your-own-key (BYOK) configurations, making it suitable for organizations with strict data governance requirements. The site notes that image, PDF, and web search inputs are supported today, with additional modalities and custom tool calling available on a per-request basis. Enterprise and self-hosted plans are tailored to individual needs and scale.

## Features
- AI Function optimization from human feedback
- Prompt engineering automation
- Batch inference via REST API
- Single/real-time inference via REST API
- Model-agnostic (open-source and proprietary models)
- Automatic model selection
- Image, PDF, and web search input support
- Self-hosted and SaaS deployment options
- BYOC and BYOK support
- Annotation queue for low-confidence results
- Continuous function improvement over time
- Classifier, judge, extractor, resolver, and router function types

## Integrations
REST API, Custom model subscriptions, Open-source models, Proprietary LLMs

## Platforms
WEB, API

## Pricing
Paid

## Links
- Website: https://sutro.sh
- Documentation: https://docs.sutro.sh/
- EveryDev.ai: https://www.everydev.ai/tools/sutro