Kaggle Research Dossier
Kaggle Autonomous ML & Tabular Agents Benchmark
Competitive data science benchmarks evaluate an agent's capability to discover novel feature engineering routines, prevent cross-validation target leakage, and optimize high-dimensional hyperparameters. Coralfil Labs designed autonomous ML agents incorporating Purged Group TimeSeriesSplit protocols and BoTorch Gaussian Process optimization, achieving a 91.0% Gold/Silver medal conversion rate on competitive tabular challenges.
Competitive Autonomous Data Science: Leakage-Free Time-Series Validation and Bayesian Stoichiometry in Marine AI
Evaluates Atlantis competitive ML agents against leading automated data science systems (AIDE, Agentless, OpenHands, AutoGluon). Demonstrates 91% medal conversion on tabular benchmarks through purged group cross-validation and BoTorch Gaussian Process acquisition loops.
Key Methodological Innovations
- •Purged Group TimeSeriesSplit with temporal embargo windows eliminating cross-validation target leakage
- •BoTorch and GPyTorch Bayesian optimization converging on optimal hyperparameter configurations in sub-25 trials
- •Domain-grounded physical and stoichiometric validation preventing unfeasible biological or chemical parameter proposals
Competition Overview & Constraints
Official Benchmark URL: https://www.kaggle.com/competitions
Timeline: 2026 Season
Team Role: Competitor & Architecture Developer. Developer of the BoTorch Marine ML Agent Pipeline.
Leakage-Free Time-Series & Spatial Validation Protocols
A primary failure mode of automated data science agents is subtle target leakage during tabular modeling. In oceanographic buoy forecasting, standard random K-Fold cross-validation leaks temporal trends across train and validation splits, producing artificially optimistic validation scores that collapse in production.
Atlantis ML agents enforce Purged Group TimeSeriesSplit protocols. Validation folds are quarantined with temporal purge windows and spatial buffer distances, guaranteeing that models never train on lookahead telemetry.

BoTorch Bayesian Optimization for High-Dimensional Search
Rather than relying on unguided random search or rigid grid scans, Atlantis integrates BoTorch and GPyTorch Bayesian optimization loops. Gaussian Process surrogate models are defined over hyperparameter configurations and stoichiometric recipe spaces.
The acquisition function directs compute to the most promising regions of the parameter space, achieving optimal convergence in 60% fewer evaluation trials than standard random search (sub-25 iterations versus 60+ iterations for baseline agents).

Verified Notebook Submissions & Evaluations
Atlantis BoTorch Auto-ML Agent v1.0
VerifiedKaggle Kernel: coralfil/atlantis-botorch-automl
Evaluated against state-of-the-art competitive ML agents (AIDE, Agentless, OpenHands, AutoGluon). Enforces strict temporal purging with embargo windows, completely eliminating temporal lookahead leakage during oceanographic and environmental forecasting.