Chasing F1 scores on a twin-turbo server

We got a dedicated Parabricks server from Nvidia.

It is nicknamed twin turbo because of the two massive enterprise GPUs humminig inside. We wanted to see if we could use it for both local LLM inference (switching v1 from API calls to a local 12B model running via vLLM and NeMo) and raw secondary analysis.

The first thing you notice when you start a job on this server is the noise. It is incredibly loud—loud enough that we only run heavy analysis pipelines at night when the office is empty.

The pipeline rabbit hole

I spent the last month going deep into secondary analysis, trying to build a reliable bio-pipeline from scratch. My goal was simple: take raw FASTQ sequencing files and output high-confidence variant calls (VCF).

It turns out “calling variants” is not a single command. It is a highly sensitive chain of tools, and every parameter tweak alters your precision and recall. I spent weeks chasing a better F1 score.

The flow I settled on:

1. Alignment: Aligning raw reads to the reference genome using Parabricks’ highly optimized FQ2BAM (which runs in minutes instead of the hours standard BWA takes on standard CPUs).
2. Variant Calling: Calling somatic mutations with Mutect2 and applying standard GATK filters, and comparing it against Nvidia’s deep-learning caller, DeepSomatic.
3. Benchmarking: Running truth sets to evaluate which combinations yielded the fewest false positives without missing critical clinical targets.

Automating the run with Agent Skills

Since these runs take a while (and you don’t want to manually baby-sit bash scripts at 2 AM), we started playing with the agent wave. We deployed Hermes directly on the server.

I wrote a set of custom agent skills to monitor the jobs. Instead of polling terminal windows, the agent is granted access to:

• Trigger FQ2BAM or DeepSomatic runs.
• Read the intermediate logs to check resource allocation.
• Parse the final accuracy metrics once the pipeline outputs the VCF.

The outcomes looked incredibly promising. For the first time, we had a bridge between a language model and a high-performance compute pipeline. But as cool as it is to have an AI coordinate a multi-hour bash pipeline, it is also a reminder that some tasks are better off as hardcoded, deterministic orchestrations.