---
title: Production ML case studies by Sudhanva Narayana
description: Public scope, engineering decisions, tradeoffs, reliability, and measured outcomes.
canonical: https://sudhanva.me/work/
last-updated: 2026-10-01
---

# Production ML case studies by Sudhanva Narayana

Sudhanva Narayana is a Senior Machine Learning Engineer specializing in
production ML platforms, high-throughput inference, and distributed model
serving. These first-party case studies document the architecture decisions,
operational tradeoffs, and measured outcomes behind his work. Proprietary
employer details and scientific data are omitted (the PyTorch DDP study is
independent research).

## Case studies

- [Billion-Row Batch Inference](https://sudhanva.me/work/case-studies/billion-row-batch-inference/):
  more than ten billion rows scored end to end in under three hours.
- [Production ML Platform on Kubernetes](https://sudhanva.me/work/case-studies/production-ml-platform-kubernetes/):
  Ray and Flyte orchestration, observability, internal scientific tooling, and
  more than $500,000 in annual cloud-cost reduction.
- [Multi-GPU Kubernetes Delivery Platform](https://sudhanva.me/work/case-studies/multi-gpu-model-build-deployment/):
  a multi-GPU Kubernetes cluster with CI/CD that doubled model release velocity
  by replacing manual build, tuning, and deploy loops with an automated
  pipeline.
- [Multi-GPU LoRA Fine-Tuning of Gemma 4 with PyTorch DDP](https://sudhanva.me/work/case-studies/pytorch-ddp-gemma-4-fine-tuning/):
  an independent engineering study with a standalone `torchrun` training
  script, 65.6M trainable LoRA parameters at rank 16, and a 35–40 GB per-GPU
  memory footprint for Gemma 4 12B at 2,048 tokens.
- [Multi-Regional Geospatial ML](https://sudhanva.me/work/case-studies/multi-regional-geospatial-ml/):
  more than 1 TB of imagery processed daily, 75% higher inference efficiency,
  and 50% lower map-rendering latency.

The real-time transformer pipeline at Autodesk, which scored more than one
million daily interactions for next-click prediction, is covered in a dedicated
[engineering post](https://sudhanva.me/blog/posts/real-time-transformer-prediction-at-autodesk/).

The [human-readable work index](https://sudhanva.me/work/) also features
open-source infrastructure projects and links to related technical writing.