Amanpreet
Singh Saimbhi

01Domains

1. 01Medical imaging

   Medical images do not arrive as images. They arrive as DICOM: pixel data wrapped in hundreds of tags describing the patient, the study, the series, the scanner and every setting it was running. Most of the difficulty lives in those tags, in routing studies to the right place, reconciling fields that each vendor fills in slightly differently, and keeping a series intact as it moves. My time mostly goes into striking a balance between what the data will support and what the clinic can absorb.

2. 02Machine learning

   I have pointed roughly the same set of methods at scans, at video, at compressed images, and at signals coming off a wrist, often several of them at once, where the useful part is what one modality knows that the others do not. The domain changes a lot and the work changes very little: find the structure, get it working outside a notebook, and be explicit about what it cannot tell you. That last one is where most of the judgement sits.

3. 03Software engineering

   I write code assuming I will not be the one running it in three years. In practice that means tests that fail for a reason, a build somebody else can trigger without asking me, and boring choices anywhere a clever one would age badly.

4. 04Software security

   Source code is a graph well before it is text. Looking at it that way turns finding a vulnerability into pattern recognition rather than a search for bug shapes you already know, and pattern recognition generalises where a fixed set of rules does not.

5. 05AI safety

   I care more about what a model can honestly claim than about what it scores. Some of the more useful work I have been part of showed that a promising idea did not help anyone. A negative result closes a door, and closing doors is how a field stops spending years on the wrong one.

6. 06Systems & infrastructure

   Nobody notices this layer until it breaks, which is most of why I like it. Queues that put the urgent thing first, containers that behave the same on a machine I will never see, pipelines that run without anyone watching them. Good infrastructure gets measured in problems that never happened.

02Method

Different domains, and the central questions stay the same. This is the part that transfers.

1. 01

   Find the structure.

   Whatever the data is, something in it is already organised. Scans, source code and sensor traces all reward the same kind of looking, and most of the work is deciding what to look at.

2. 02

   Get it running on real data.

   A result on a clean benchmark is a hypothesis. A result on live traffic, at the speed the people using it need, is a finding. Almost everything interesting shows up in the gap between the two.

3. 03

   Make it survive other people.

   Tests that fail for a reason, a build anyone on the team can trigger, and boring choices anywhere a clever one would age badly. If it only runs when I am in the room, it does not run.

4. 04

   Be clear about what it cannot do.

   Every model has an edge past which it is guessing. Finding that edge and saying where it is takes the most care and gets skipped the most often, and in medicine it is the part that matters most.

03Toolkit

Languages

Python · C/C++ · JavaScript · Go · SQL · Java · PHP · C#

Machine learning

PyTorch · MONAI · TensorFlow · Keras · scikit-learn · OpenCV · NumPy · pandas

Web

Flask · Django · React · Node · Express · Plotly.js · .NET

Infrastructure

Docker · AWS · Jenkins · Git · pytest · DICOM · Metabase

Data

PostgreSQL · MySQL · MSSQL · Oracle · MongoDB · SQLite

04Trajectory

1. 2023 – PresentResearch Software Engineer, NYU Grossman School of Medicine New York
2. 2022Software Engineer Intern, Next Jump New York
3. 2021 – 2023Graduate Researcher / Research Assistant, New York University New York
4. 2021 – 2023MS, Computer Science Engineering, New York University New York
5. 2017 – 2021BTech, Computer Science Engineering, Guru Gobind Singh Indraprastha University New Delhi

05Off duty

New Delhi to New York was less a change of address than a change of defaults. Almost everything I was already good at had to be relearned in a different context, and the parts nobody teaches you, how to disagree without stalling a project, when to ask instead of guessing, when a thing is actually finished, ended up mattering more than anything I already knew how to build.

Away from all of that: the gym most evenings, which is the one system I work on with a completely legible feedback loop. Travel the rest of the time, mostly as an excuse to be a beginner at something again.