status: in development · ray + jina + dask + blender on k8s

Raw data goes in. Trained models come out.

Smelt is the ML foundry for the Colossal Capital ecosystem — distributed training on Ray, neural serving on Jina, parallel feature engineering on Dask, and a headless Blender farm for rendering physics simulations into NFT-grade visuals. One cluster, every model the platform runs on.

View on GitHub

# distributed training with ray import ray from ray import train trainer = train.torch.TorchTrainer( train_loop_per_worker=train_model, scaling_config=train.ScalingConfig(num_workers=4), ) result = trainer.fit() # → fanned out across the GPU cluster

What gets forged

Six model families in production rotation

Market predictionPrice forecasting, market-cycle detection, trend analysis, and volatility prediction across crypto and traditional markets.LSTM · transformer
Portfolio optimizationPPO reinforcement-learning agents for multi-asset trading: risk-adjusted construction, dynamic rebalancing, and backtesting at scale.PPO · RL
NFT analyticsRarity scoring, valuation models, trait analysis, and market sentiment — including spotting undervalued collections.rarity · valuation
Physics simulationsMagic-square calculations, world-line and world-sheet rendering, and particle simulations, output through Blender as 3D assets.blender · physics
Trading signalsSignal generation with real-time adaptation, automated strategy backtesting, A/B testing, and performance monitoring.signals · backtest
Model managementVersion control, automated retraining, drift detection, and monitoring — the MLOps pipeline that keeps the rest honest.mlops · drift

Infrastructure

Four rigs, one Kubernetes floor

distributed compute

Ray

Training across GPU clusters, parallel hyperparameter tuning, model selection, and inference at scale.

model serving

Jina

Neural search and serving: document understanding, embedding generation, and real-time inference APIs.

data processing

Dask

Parallel feature engineering with lazy evaluation, distributed DataFrames, and streaming.

3d rendering

Blender

Headless render farm for physics NFTs — world lines and world sheets visualized at scale.

Who it feeds

Every prediction in the ecosystem traces back here

AckwardRoots streams the training data in — market feeds, chain events, financials. Delt takes predictions, signals, and portfolio recommendations into the trading app. TWF uses rarity prediction and trait analysis for the turtle-breeding ecosystem. Aether gets Blender-rendered physics NFTs. Versal displays live market predictions in the terminal, and 5th Element Ink runs its 3D asset pipeline on the same render farm.

Smelt is internal infrastructure for the CC ecosystem — there's no standalone subscription. Platform access follows your Colossal Capital tier via the keys portal.