Inspiration

Clinical genetics laboratories and academic bioinformatics groups generate massive sequencing cohorts (FASTQ, BAM, CRAM, VCF) alongside scanned patient requisition forms, wet-lab sample tube photos, bench voice notes, and post-pipeline quality and variant reports. Because human genomic data is inherently self-identifying, strict data-protection regulations (HIPAA, GDPR, and clinical data-residency policies) prohibit uploading patient sequences or clinical phenotypes to third-party cloud AI services. Scientists are forced to store cohorts on local NAS volumes and manually check cryptographic hashes, match sample tube labels, and inspect QC and variant reports one patient at a time.

What it does

Genomic Cohort Triage (genomic-cohort-triage) is a 100% local, zero-cloud Uliya orchestration Skill for UGREEN AI NAS. Once local bioinformatics pipelines finish writing outputs to a NAS cohort directory, a scientist can trigger the Skill with a single sentence to audit the entire cohort in one pass:

  1. Cohort Inventory & Cryptographic Integrity: Lists sequencing and report files, calculates sha256sum digests to catch transfer corruption, and detects accidental duplicate files via scan-dedupe.
  2. Wet-Lab Tube OCR & Gel Inspection: Uses on-device computer vision (vision-ocr and vision-vlm) to read sample IDs printed on cryotubes and inspect electrophoresis gel images locally.
  3. Multi-Source Clinical & Variant Reconciliation: Cross-checks patient informed-consent status, referral phenotypes, sequencing coverage (mean depth >= 20x, Q30 >= 80%), and ACMG Class 4/5 pathogenic variants across PDFs and tables, quoting the exact source file and page number.
  4. Bench Voice Memo Integration: Transcribes wet-lab bench audio recordings (such as sample hemolysis or extraction notes) and links them to the corresponding sample ID.
  5. Approval-Gated Cohort Filing: Displays a ranked Cohort Triage Dossier in chat first (BLOCKED_CONFLICT, MISSING_CONSENT, CRITICAL_VARIANT_ALERT, QC_FAILURE, VERIFIED_PASS). Only after explicit user confirmation and UGOS platform approval, it saves the dossier, copies verified files into structured cohort directories (preserving originals untouched), and applies status tags.

How we built it

We architected genomic-cohort-triage strictly according to the Uliya Skill Development Standard using the official UGREEN Hackathon Starter Kit (uliya-skill-kit v0.7.3) and Capability Catalog v2.0.0. The package is pure UTF-8 Markdown (SKILL.md plus four directly linked specifications under references/) with zero executable scripts, binaries, or external API calls. The parent Skill delegates self-contained tasks across four registered UGOS expert agents (file-agent, photo-agent, document-agent, and knowledge-agent), passing only verified upstream outputs forward and explicitly handling all five delegation states (completed, failed, cancelled, waiting_user, waiting_approval).

Challenges we ran into

  • Operating within Declarative Orchestration Constraints: Because Uliya Skills cannot execute arbitrary Bash scripts or trigger Docker containers directly, we designed the Skill around the post-pipeline data bottleneck—auditing and reconciling the multi-modal files (FASTQ/VCF hashes, tube photos, consent PDFs, QC summaries, and variant tables) that already converge on the NAS.
  • Zero-Hallucination Clinical Safety: In clinical genomics, a guessed sample ID or fabricated variant is unacceptable. We enforced a strict five-label provenance contract (VERIFIED_FACT, AI_INTERPRETATION, CONFLICT, MISSING, UNCLEAR) that halts automated filing whenever a sample ID conflict (SAMPLE-104B vs SAMPLE-104A) or unsigned consent form is detected.

Accomplishments that we're proud of

  • Designing a domain-specific clinical genomics and bioinformatics workflow that maps 100% to published Capability Catalog v2.0.0 actions across four UGOS expert agents.
  • Passing all static validation rules (0 errors, 0 warnings) with deterministic failure and abstention handling.

What we learned

Working with Capability Catalog v2.0.0 demonstrated how combining local file hashing (ug-filesystem), on-device OCR/VLM (ug-photo), multi-document reconciliation (ug-document), and voice transcription (ug-knowledge) turns a NAS from passive cold storage into an active, privacy-preserving laboratory compliance engine.

What's next for Genomic Cohort Triage

In Round 2, we plan to deploy and benchmark genomic-cohort-triage on the UGREEN AI NAS iDX6011 Pro (64GB) using our 12-sample synthetic genomic cohort fixture to measure multi-gigabyte sha256sum throughput, curved-tube vision-ocr accuracy on the 96 TOPS NPU, and live UGOS approval prompts.

Built With

  • bioinformatics
  • file-workflow
  • local-ai
  • markdown
  • ocr
  • ug-document
  • ug-filemgr
  • ug-filesystem
  • ug-knowledge
  • ug-photo
  • ugos
  • ugreen-ai-nas
  • uliya
  • uliya-skill-kit
Share this project:

Updates

Submission history