A*STAR, Singapore - GNOME Dark Matter Detection

This one is hard to summarize without first explaining what the project actually is, because it's genuinely unlike anything else I've worked on.

The Science

The GNOME Project is a global physics collaboration spanning 9 to 12 countries. Each country runs a "station" equipped with magnetometers that function as sensors, and together this network listens for signals of dark matter and other exotic physics phenomena. The idea being that if something like dark matter passes through Earth, it would leave a correlated signal across multiple geographically separated sensors at roughly the same time.

That's the physics. My job is the data side of it.

The Data Processing

Here's the problem with a network like this: every station sits in a different country, on a different power grid, in a different time zone. The raw data needs to be synced to a global GPS timestamp before anything useful can be done with it. On top of that, each country's grid introduces power-line noise at either 50Hz or 60Hz depending on where the station is, and the sensors themselves are calibrated every hour, meaning any data collected during an uncalibrated window is unreliable and needs to be thrown out.

Before I joined, none of this had a unified solution. It was all manual.

There was no single preprocessor that could handle the full suite of stations across a consistent pipeline. So I built one.

Then I ran it across 6 months of raw data spanning their entire station network, which meant profiling and processing 347,000 raw HDF5 files across 9 stations and building the multi-terabyte ELT pipelines and AWS data lake workflows to handle it at that scale.

To validate the output, I reproduced the project's SR5 and SR6 reference outputs to greater than 99% fidelity, which was the benchmark for confirming the preprocessor was actually doing what it was supposed to. I also ran an LSTM baseline on T4 GPU compute to probe whether anomaly detection over this data was even feasible, which it is, and that opens up a lot of what comes next for the project.

The Full-Stack Development Story

The other side of the work was access. Physicists across the collaboration needed a way to actually get at this processed data for their theoretical work and experiments. So I built a portal that lets them pull processed data from that 6-month window across whatever combination of stations and days they need, packaged as a zip file ready to go. It also has a built-in forum for centralized discussions across the collaboration so that conversations about the data, the analysis, and the science aren't scattered across emails and Teams threads.

Somewhere between building ETL pipelines for terabytes of physics data and thinking about what dark matter signals might actually look like in a time series, this stopped feeling like an internship and started feeling like real work.

I'm designing the Singapore-side cloud and ML stack for a project that is, at its core, trying to detect dark matter. It's a strange and excellent thing to be doing.

Been here since January 2026, who knows what the future holds. And I'm excited about it.

as a bonus, here's a picture of me from the day I completed my first week at A*STAR :)