Systems That Last.

Strength doesn't need to be loud. It needs to be steady.

Can Triton Fit It All?

One server, multiple frameworks, zero apologies for the simplification.

An intro to NVIDIA Triton Inference Server, why decoupling model inference from the serving flow solves a real resource and versioning problem, and where Triton stops making sense.

Posted by Varchaleswari Ganugapati on Sunday, July 5, 2026

The Handshake: Validating Score Parity Across Training and Serving

Part 2. The model traveled. But are the scores actually equal?

Part 2 of cross-platform model bundling. How to prove that a model trained in Python produces identical scores when served from Java: golden records, shadow scoring, statistical tolerance bands, and case studies from Uber, LinkedIn, Airbnb, DoorDash, and Netflix.

Posted by Varchaleswari Ganugapati on Saturday, May 16, 2026

Part 1: Cross-Platform Model Bundling: Formats That Travel

Trained in Python. Served in Java. Scored in milliseconds. The format in between is the whole story.

A deep look at cross-platform model bundling formats: MLeap, ONNX, PMML, MLflow, TorchScript, and BentoML. Not a format survey but a framework for choosing, with real-world examples from Meta, Microsoft, Hugging Face, and Databricks.

Posted by Varchaleswari Ganugapati on Friday, May 15, 2026

Powering Models: Feature Engineering, Part 2

Getting features to the model at inference time: passing them in the request or fetching them from a store. Both choices have consequences.

Part 2 of 2. How production services make features available to models at inference time: the tradeoffs between request-time feature passing and online feature store retrieval, with patterns from real systems.

Posted by Varchaleswari Ganugapati on Wednesday, May 13, 2026

Powering Models: Feature Engineering, Part 1

From raw events to training-ready features: the offline pipeline that makes production models possible.

Part 1 of 2. A deep dive into the offline half of the ML lifecycle: EDA, the DS to engineering handoff, feature pipelines on Kafka and GCS, and automating it all with orchestration tools.

Posted by Varchaleswari Ganugapati on Wednesday, May 13, 2026