Machine Learning

Production machine learning, engineered — not notebooks.

Plenty of teams can train a model. The hard part is what comes next: distributing billions of calculations across Kubernetes, serving models behind real APIs, and turning a data scientist’s research prototype into a hardened, unit-tested production system. That’s the part Leopard Data does — distributed training and inference on Kubernetes with Ray and KEDA, deep learning in PyTorch, and deterministic ML.NET inside enterprise .NET.

Start a Project See the Proof

Shoulder-to-Shoulder with Data Science

We take the math to production scale.

Our role on ML engagements is the bridge most teams are missing: we work hand-in-hand with data science teams, take the models they’ve proven in research, and engineer them into distributed, tested, monitored production systems — 30+ years of enterprise engineering applied to feature sets of 200K+ features and workloads of hundreds of millions to billions of calculations. The data scientists own the math; we make it finish, ship, and scale.

What We Do in Machine Learning

Six disciplines, all backed by shipped production systems.

Distributed ML on Kubernetes

When a model has to run against hundreds of millions to billions of calculations, one machine won’t do. We parallelize ML workloads across Kubernetes with KEDA event-driven autoscaling and Ray clusters on AWS EKS — scaling workers to the queue, not to a guess.

Deep Learning in Production

LSTM networks in PyTorch and DeepAR forecasting — integrated into ensembles alongside classical methods like ARIMA, Holt-Winters, and Prophet, so the deep-learning model earns its place against the baselines instead of being assumed.

Feature Engineering at Scale

Ranking engines that evaluate data-science models against feature sets of 200K+ features — with the distribution strategy chosen by benchmark, not fashion: we’ve compared Ray/Anyscale, KEDA, and Azure Durable Functions head-to-head on the same workload. See the pipelines that feed the models

ML in .NET

ML.NET time-series forecasting (SSA) and SDCA regression running inside enterprise .NET applications — deterministic, unit-testable, production-grade ML for teams that need repeatable outputs, not a Python sidecar.

MLOps & Model Serving

Models don’t ship themselves. We integrate them behind REST APIs with job tracking and monitoring, serve and train on SageMaker where it fits, and wire the surrounding platform — including Kong middleware integration — so ML behaves like any other production service.

Partnering with Data Science

We turn research prototypes into hardened production systems — full unit and integration testing around the math, reproducible pipelines, and engineering review the data-science team can build against. For the LLM and agent side of the practice, see AI-first development

Machine Learning in the Wild — Real Engagements

Distributed ML for industry, deep-learning forecasting, and deterministic ML.NET in a shipped SaaS.

Industrial · Distributed ML

Koch Industries — Ranking & Prediction at Billion-Calculation Scale

A ranking solution running data-science math models against feature sets of up to 200K features — hundreds of millions to billions of calculations — distributed across Kubernetes with KEDA on AWS EKS, chosen after hands-on POCs (RabbitMQ + minikube first, then EKS) and benchmarked against an Anyscale/Ray prototype. Plus a prediction engine: an LSTM algorithm in PyTorch on Ray.io for parallelism, designed and integrated hand-in-hand with the data scientists, with full unit and integration testing.

Read the full case study

Industrial · Forecasting Ensembles

Distributed Forecasting Ensembles on EKS

Production-capacity forecasting for an industrial forecasting platform — ensemble models combining ARIMA, Holt-Winters, Prophet, LSTM, and DeepAR, served behind REST services and parallelized on EKS with KEDA so forecasts across business units actually finish on schedule.

Read the full case study

FinTech SaaS · ML.NET

Grade My Investments — Deterministic ML.NET in Production

ML.NET SSA time-series forecasting and SDCA regression grading US stocks deterministically — backed by 596 tests — with Claude as a language layer on top. The pattern worth copying: a deterministic ML core doing the math, an AI explanation layer doing the talking.

See the architecture

Have models that need to reach production scale?

From distributed training and inference on Kubernetes to deterministic ML.NET inside enterprise .NET, Leopard Data engineers the machine learning that ships. Corp-to-Corp engagements out of Plano, TX.

Start a Project See All Services