Siddharth Rao

Machine Learning Engineer at Morrowleaf Health | Building reliable models for clinical workflows

I’m a machine learning engineer at Morrowleaf Health in Bengaluru, working on models that help care teams prioritise follow-up without adding another noisy dashboard to their day. My work spans evaluation, deployment and the less glamorous business of finding out why something behaved differently on Tuesday. Outside work, I run trails and occasional 50ks, build mechanical keyboards, and practise Carnatic violin—usually before the neighbourhood is fully awake. I like careful work, long climbs, and making things a little less complicated for the next person.

Build and maintain models that support care-team follow-up prioritisation. Own evaluation pipelines, production monitoring and error analysis with clinical operations colleagues. Introduced patient-level dataset splits and subgroup reporting into release reviews; mentor two engineers on deployment and debugging.

Developed forecasting and anomaly-detection systems for operational teams. Took models from exploratory notebooks to monitored services, with particular attention to reproducible training and clear fallback behaviour.

Built Python services and data-processing jobs for a workflow software product. Gradually moved into applied machine learning after spending enough time debugging the data feeding it.

Python, PyTorch, SQL, Machine Learning, Model Evaluation, MLOps, Time-Series Modelling, Data Pipelines, Production Monitoring, Technical Mentoring

English, Kannada, Telugu

A useful question from our clinical operations review this week: “What should I do differently because of this score?” We had a good explanation of how the model worked. We had a less good explanation of what happened next. Spent the afternoon tightening the handoff, including when to ignore the recommendation and where to record that decision. A model can be technically sound and still leave someone with extra work. That is also a bug.

One of our newer engineers asked why I insist on keeping a small, manually reviewed evaluation set alongside the larger automated one. Today we found an upstream change that made a missing value look like a genuine zero. The aggregate metrics barely moved. The individual examples made the problem obvious. No dramatic lesson here: keep a few examples you understand very well, and look at them when the pipeline changes. Boring checks earn their place.

Finished a 50k near Sakleshpur on Sunday. My carefully constructed pacing plan survived approximately one enthusiastic downhill. The rest was slower, wetter, and helped considerably by volunteers who kept offering food when I had forgotten that eating was part of the job. Back at my desk today, with a healthy respect for stairs. Grateful to the organisers, the aid-station teams, and my friend who waited at the finish with dry socks. Excellent project management.

We changed our model review template this month. Before discussing headline performance, we now look at the most consequential misses, the groups with thin evaluation data, and what the fallback does. It makes the meeting slightly longer and the release decision much clearer. Working in healthtech keeps reminding me that an average is a useful summary, not a complete account of what someone will experience. Credit to our clinical colleagues for asking the uncomfortable questions early.