Model & pipeline review
A senior audit of your models and training stack — architecture, tf.data pipelines, training stability, evaluation methodology, and serving path — with findings tied to measured improvement targets.
Applied model development
From problem framing and dataset curation through transfer learning, fine-tuning, and hyperparameter methodology. Training runs that are documented, reproducible, and honest about their evaluation.
Keras 3 & backend modernization
Migrate legacy TF1.x and tf.keras codebases to Keras 3, gain backend portability across TensorFlow, JAX, and PyTorch, and pick up the performance work: mixed precision, distribution strategies.
Deployment & MLOps
The road from checkpoint to production: TF Serving, ONNX, and LiteRT export for edge, model versioning, drift monitoring, and retraining pipelines that keep shipped models trustworthy.
Baselines before novelty
We start every modeling problem with data quality and a strong baseline, because most applied-ML wins live there — not in exotic architectures. Improvements are claimed only against a measured reference.
Multi-backend depth
Keras 3 runs on TensorFlow, JAX, and PyTorch. We work across all three backends — including the stateless training style JAX demands — so your code targets the runtime that fits, not the one it inherited.
The notebook is not the deliverable
An engagement ends with a serving path: exported artifacts, parity tests, monitoring, and a retraining route. We build the parts that keep working after we leave.