---
title: Production ML systems expertise | Sudhanva Narayana
description: Evidence-backed expertise in ML platforms, inference, serving, and reliability.
canonical: https://sudhanva.me/expertise/production-ml-systems/
last-updated: 2026-10-01
---

# Production ML systems expertise | Sudhanva Narayana

At Montai Therapeutics, Autodesk, and Pixxel, Sudhanva Narayana has built and
run the systems behind production machine learning, including batch pipelines scoring 10B+
molecules in under three hours, real-time transformers serving 1M+ daily user
events, multi-region satellite imagery pipelines, and the Kubernetes, Ray, and
Flyte clusters underneath them.

## Capabilities

- Shared ML compute platforms on Kubernetes, Ray, and Flyte, provisioned with
  Terraform and Argo CD, with MLflow tracking and model registries.
- Distributed batch prediction (10B+ rows in under three hours) and low-latency
  online model serving on multi-GPU clusters.
- Multi-GPU training and LoRA fine-tuning in PyTorch with DDP, `torchrun`, and
  NCCL, grounded in explicit GPU memory and bandwidth math.
- Prometheus and Grafana monitoring, alert rules, ingestion schema validation,
  and GPU node autoscaling that cut annual cloud spend by over $500K.
- Data-heavy and scientific ML pipelines, including molecular screening, 1
  TB+/day hyperspectral satellite imagery, and cross-AZ Milvus vector search.

## Evidence

- More than 10B rows scored by a distributed TensorFlow batch inference pipeline
  in under three hours.
- More than $500K in annual cloud-cost savings across production ML
  infrastructure.
- 2x model release velocity and lower GPU cost per release from a multi-GPU
  Kubernetes delivery platform with automated build, tuning, and deployment.
- More than 1M daily user interactions scored in real time by a Transformer
  serving pipeline at Autodesk
  ([engineering write-up](https://sudhanva.me/blog/posts/real-time-transformer-prediction-at-autodesk/)).
- A complete PyTorch DDP training architecture for LoRA fine-tuning Gemma 4 12B
  across multiple GPUs
  ([engineering study](https://sudhanva.me/work/case-studies/pytorch-ddp-gemma-4-fine-tuning/)).

- [Complete expertise page](https://sudhanva.me/expertise/production-ml-systems/)
- [Production case studies](https://sudhanva.me/work/)
- [Technical writing](https://sudhanva.me/blog/)