Daft is a declarative query engine that enables engineers to write SQL‑like queries for multimodal data—images, video, audio, and text—automatically handling loading, preprocessing, and GPU‑accelerated distributed execution. It offers a unified schema‑on‑read data model, automatic partitioning, fault tolerance, and secure role‑based API access, with native connectors and Python/Rust SDKs for seamless integration into existing ML pipelines.
Funding
Funding not disclosed
Founders
Product
Problem
AI teams must ingest and analyze massive volumes of images, video, audio, and text, but existing pipelines require deep expertise in distributed systems and custom engineering. This complexity leads to prolonged development cycles and limits the ability to prototype multimodal models quickly. Additionally, the lack of a unified query interface forces engineers to stitch together disparate tools, increasing operational overhead.
Solution
Daft is a declarative query engine that treats multimodal data with the same simplicity as SQL handles tabular data. Engineers write high‑level queries that automatically orchestrate data loading, preprocessing, and parallel execution across clusters without writing low‑level distributed code. The platform provides a unified data model and native support for GPU‑accelerated operators, enabling petabyte‑scale processing of images, video, audio, and text. Built‑in fault tolerance, automatic data partitioning, and schema‑on‑read capabilities let teams focus on model development rather than infrastructure. Daft integrates with major cloud storage services and exposes Python and Rust SDKs for seamless embedding into existing ML pipelines. All query results are delivered through a secure, role‑based API that supports enterprise compliance standards.
Target Audience
The primary customers are data‑science and ML engineering teams at enterprises that need to process petabyte‑scale multimodal datasets—such as autonomous‑vehicle developers, large e‑commerce platforms, and media analytics firms. Daft also serves AI research groups building next‑generation multimodal models that require rapid, reliable data pipelines.
Features
- Declarative multimodal query language with SQL‑like syntax for images, video, audio, and text
- Unified data model that abstracts storage formats and enables schema‑on‑read processing
- Automatic data partitioning and distributed execution engine optimized for GPU clusters
- Built‑in fault tolerance, retry logic, and progressive result streaming for large datasets
- Native connectors to S3, GCS, Azure Blob, and on‑premise object stores with zero‑copy ingestion
- Python and Rust SDKs plus RESTful API for integration into existing ML workflows
- Role‑based access control and end‑to‑end encryption to meet SOC 2 and GDPR requirements
- Integrated data cleaning primitives (deduplication, format conversion, metadata extraction) that run in‑pipeline