VEOX
VEOX builds engineered intelligence systems that search for novel model structures rather than tuning parameters of known algorithms. Its first product, jAIn, jointly optimizes preprocessing, model family, hyperparameters, ensembling, pruning, and distillation while scoring candidates on both accuracy and serving cost. The output is plain code and weights the customer owns and can deploy anywhere, with full run logs published for auditability.
- Artificial Intelligence
- Developer Tools
- Software Only
Funding
Founders
Product
Problem
Traditional machine learning tools require humans to specify model architectures, hyperparameters, and feature engineering steps, limiting discovery to structures someone already thought of. This approach misses novel compositions that could perform better on both accuracy and serving cost, and it often produces models that are expensive to run in production.
Solution
VEOX builds engineered intelligence systems that search a space of model structures rather than fitting members of a named family. Its synthesis engine, RecurEvolv, jointly searches preprocessing, model family, hyperparameters, ensembling, pruning, and distillation as one integrated process. The first product, jAIn, scores every candidate on both held-out accuracy and inference cost, ensuring lean models that are practical to serve. What comes back is plain code and weights the customer owns and can deploy anywhere, along with a full log of what was tried and rejected. The company publishes its methodology and results protocol in advance, including failed runs, to maintain transparency.
Target Audience
Primary customers are data science and machine learning teams at enterprises that need to push beyond conventional model performance and want full ownership of their deployed models.
Features
- Searches model structures rather than tuning parameters of known models, enabling discovery of novel architectures that have no name yet
- Jointly optimizes preprocessing, model family, hyperparameters, ensembling, pruning, and distillation in one search rather than six sequential decisions
- Scores every candidate on both held-out accuracy and serving cost, favoring lean models over complex ones
- Returns deployable code and weights the customer owns and can run anywhere, not an API endpoint
- Publishes full run logs including rejected candidates for auditability
- Includes a documented protocol for measuring results before results exist, with the research preview delivering receipts rather than summary numbers