Contig
Agentic bioinformatics: it runs, self-heals, and verifies
7 followers
Agentic bioinformatics: it runs, self-heals, and verifies
7 followers
An agentic analyst for sequencing data: it runs the right pipeline on your own compute, self-heals failures like out-of-memory and missing indexes, verifies the output with real QC checks, and returns a signed, reproducible record. CLI + local dashboard.
how does the reproducibility work if my compute or reference versions drift between runs, is the signed record still useful when the underlying environment changes?
@berrazwqi Great question, and it gets at something we thought hard about. The signature isn't a promise that your environment stays frozen forever, because it can't be. It's a tamper-evident record of exactly what produced a given result: pipeline revision, resolved parameters, container digests, tool versions, input and output checksums, and the reference identity (the FASTA/GTF hash or iGenomes key). The Ed25519 signature just guarantees nobody edited that record after the fact.
So when your compute or reference versions drift, the signed record actually becomes more useful, not less. You rerun, you get a second signed record, and contig diff shows you precisely what changed: which container digest moved, which reference hash differs, and how the QC numbers shifted as a result. Because both records are signed, that comparison is trustworthy. You're not left arguing about whether the old run "really" used version X, it's pinned right there.
The record also tells you honestly whether a run is fully pinned or not, so you know upfront how reproducible it really is rather than finding out later. Short version: we don't prevent drift, we make it visible, attributable, and provable. That's usually the part that's missing.
ran contig on a messy fastq set and the self-heal caught an oom mid-run without me babysitting it, which honestly saved me an evening. love that the output comes back with the actual qc numbers baked in.
@mcahit321044 Thank you, this is exactly the moment we built Contig for. That OOM catch is the whole thesis in action. The self-heal actually reads the failed task's peak memory from the run's own trace and sizes the retry to fit, instead of blindly doubling and hoping. So it isn't just "retry bigger," it's "retry by this much, because that's what the run told us it needed," and every step gets recorded in the repair chain if you ever want to see what it did while you were away.
And yes, the QC numbers coming back inside the result instead of in some separate report you have to go dig up is very much on purpose. The point is a verified answer, not a pile of logs. Really glad it saved you the evening. If you ever hit a failure it can't self-heal, that's the most valuable thing you could send our way. 🙏