I’m Amogh. I was at GalaxEye Space from January to June, working on SAR processing and the infrastructure around it. Before that, two summers of computer vision at AWIROS. Right now I’m looking for a job where the work is systems and performance.
PulseRiver, end to end
4 GB dataset benchmark, one run
Most of it came from moving the per-stage work onto the GPU and running the stages in parallel rather than one after another. It stayed at six seconds after we added Global Back Projection, DEM support and a GBP kernel.
Three things have my evenings.
Keeping the vLLM deployment below healthy, and working out GPU scheduling under load that isn’t synthetic.
Five boxes on a 200GbE fabric. I want to know where tensor parallelism stops being worth it and pipeline parallelism takes over.
Coming from C, so ownership made sense quickly and traits took a while. Reading gitoxide and the Wild linker to find something small I can actually fix.
Most recent first.
They can’t send prompts to a hosted API, so everything runs in their cluster. I set up vLLM on Kubernetes, sorted out the GPU scheduling, and got model rollouts working without dropping requests that were already in flight. Prometheus, Grafana and Langfuse sit on top so that when latency moves we can say why.
Selling hardware here means dealing with BIS, WPC, TEC and MeitY. The rules are spread across a lot of PDFs that cite each other, and finding out whether something applies to you takes hours. I built a system that reads the whole corpus and answers questions about it, with the clause it used attached to every answer, so you can check it rather than take its word.
Radar imaging produces more data than a satellite can send down, so a lot of the processing has to happen quickly. I worked on PulseRiver, the end-to-end pipeline. Most of my time went on the processing speed, and then on the things around it that were also slow: the CI, the images, the database, and the fact that nobody could see what production was doing.
Benchmarked vision-language models for scene retrieval, where you describe what you want in plain text and get the right frames back. I built the pipelines that connect a text query to what’s actually in the pixels, which we used for automatic annotation. I also improved the OCR and crowd-counting models by self-training them on unlabelled footage, and helped get everything onto Jetson boards with TensorRT.
Detection and tracking models for live video, and the alerting modules on top of them. Everything had to run under 100 ms per frame.
Two I built on my own time.
Detection models need labelled images and there is never enough of them. I trained latent diffusion models to generate more, with CLIP conditioning to keep the prompt and the image related. Then I checked whether it was actually useful, which is the part people skip.
Aircraft broadcast their position and you can pick it up with cheap hardware. I built the receiver pipeline, the real-time processing behind it and the interface, then spent most of the time keeping it accurate while people were using it.
I put this together so I could test scheduling and parallelism myself rather than read about it. Three Grace-Blackwell boxes and two x86 servers on the same fabric.
Written down so I stop working them out twice.
Where the time went in the SAR pipeline, and which of the fixes actually mattered.
How a nineteen minute build came down to three.
What to instrument first when eight services have no traces and everyone is guessing.
Coming from C: what a fat pointer is, and when monomorphisation costs more than a vtable.
Inference infrastructure, GPU systems, performance work. I’m in Bengaluru and happy to relocate. Email is the fastest way to reach me.
Email me