Datology AI

DatologyAI, Inc. develops software that automates the curation of datasets used to train AI models, including large language models and other generative systems. The platform identifies which data in a corpus is most valuable for a given model application, removes redundant or harmful samples, suggests augmentation with additional or synthetic data, and determines how data should be batched during training. According to the company and press coverage, it can process datasets at petabyte scale and is modality-agnostic, handling text, images, video, audio, tabular data and more specialized formats such as genomic and geospatial data [2][3][4].

The company positions its product as a four-stage "data refinery": cleaning (removing poorly formatted, empty, short or evaluation-contaminated records), curating (filtering by quality, task and taxonomy), creating (generating synthetic data for task relevance and diversity), and composing (mixing sources into staged training datasets). Outputs are intended for mid-training on open models or pre-training of custom foundation models, and deployments run inside the customer's own private cloud or on-premises infrastructure. The company states it does not sell or source data itself, only processes customer-supplied proprietary, public web, and licensed data [0][2][3][4].

Published customer results include a legal-domain mid-training engagement reporting a 5% improvement on LegalBench and 2.5% on general evaluations using 100B mid-training tokens against a 15T-token pre-trained base, and a frontier-class open-weights mixture-of-experts model (398B total parameters, 13B active) trained on roughly 17-20T tokens curated by Datology, which the company says served 3.37T tokens on OpenRouter in its first two months [0].

Founding story

Ari Morcos, who holds a PhD in neuroscience from Harvard, spent two years at DeepMind applying neuroscience-inspired techniques to understand and improve AI models and more than five years at Meta's AI lab studying the mechanisms underlying model behavior, work that largely involved manipulating training data. He founded DatologyAI to abstract away data preparation work in model training, together with co-founders Matthew Leavitt and Bogdan Gaza, a former engineering lead at Amazon and later Twitter [2][3].

Business model

Enterprise software sold to organizations that train their own AI models; the product is deployed into the customer's infrastructure (on-premises or virtual private cloud) rather than requiring data to be transferred to the vendor. The company states it does not sell data or tokens, only tooling that processes customer data [0][2][3].

Not specified in the available sources.

Traction

Reported customer engagements include a legal-domain mid-training project with Thomson Reuters-referenced evaluation gains and a partnership with Arcee AI on model building; the site cites a frontier open-weights MoE model trained on Datology-curated tokens that served 3.37T tokens on OpenRouter in its first two months. Total disclosed funding reported at $57.5 million as of May 2024 [0][6].

Latest developments

The company website lists research updates dated 2026, including work on automating data research (DataSmith), inducing concision in vision-language models via data curation, and pretraining with finetuning data, alongside case studies on legal-domain mid-training and a frontier open-weights MoE pre-training engagement [0].