Skip to content

Precision Intelligence Network

Stage 2 of the HCLS AI Factory Pipeline

The Precision Intelligence Network is a coordinated system of 11 domain-specialized intelligence agents, each built on RAG (Retrieval-Augmented Generation) architecture with Milvus vector search and Claude AI. All 11 agents share read-only access to the genomic_evidence collection (3.5M vectors) produced by the Genomic Foundation Engine, while maintaining their own domain-specific vector collections.

What It Does

Raw genomic variants from Stage 1 become clinically actionable intelligence through domain-specific interpretation, cross-agent coordination, and evidence-grounded reasoning. Each agent transforms the shared molecular foundation into specialized insights for its clinical domain.

The 11 Intelligence Agents

Agent Domain UI Port API Port Collections Key Capabilities
Precision Oncology Cancer 8526 8527 11 MTB packets, therapy ranking, trial matching, FHIR R4
CAR-T Intelligence Cell Therapy 8521 -- 11 Cross-collection evidence, comparative analysis, deep research
Imaging Intelligence Medical Imaging 8525 8524 10 VISTA-3D, MAISI, VILA-M3 workflows, FHIR R4 export
Precision Biomarker Biomarkers 8528 8529 14 PhenoAge/GrimAge, 9-domain risk, genotype-aware interpretation
Pharmacogenomics Drug-Gene 8507 8107 15 Star allele calling, CPIC guidelines, 9 dosing algorithms
Precision Autoimmune Autoimmune 8531 8532 14 Autoantibody interpretation, HLA analysis, flare prediction
Neurology Intelligence Neurology 8529 8528 14 Stroke triage, dementia evaluation, EDSS scoring
Cardiology Intelligence Cardiovascular 8536 8126 13 11 clinical workflows, 6 risk calculators
Clinical Trial Intelligence Clinical Trials 8128 8538 14 Trial optimization, adaptive design, biomarker strategy
Rare Disease Diagnostic Rare Disease 8544 8134 14 HPO matching, ACMG classification, gene therapy tracking
Single-Cell Intelligence Single-Cell 8130 8540 12 Cell type annotation, TME profiling, drug response

Shared Molecular Foundation

All 11 agents connect to the genomic_evidence collection as a read-only data source. This shared foundation ensures:

  • Consistency -- Every agent reasons from the same molecular data
  • Efficiency -- 3.5M vectors are indexed once, queried by all
  • Traceability -- Variant-level lineage from FASTQ through every agent's output

Cross-Agent Coordination

Agents communicate through Server-Sent Events (SSE) and cross-modal triggers:

  • Imaging → Genomics: Lung-RADS 4A+ triggers EGFR/ALK/ROS1/KRAS genomic queries
  • Oncology → Drug Discovery: Actionable variants trigger molecule generation in Stage 3
  • Biomarker → Pharmacogenomics: Genotype findings trigger drug interaction checks
  • Autoimmune → Biomarker: HLA associations trigger biomarker monitoring recommendations

Shared Technology Stack

Component Technology
Vector Database Milvus 2.4 with IVF_FLAT/COSINE indexes
Embeddings BGE-small-en-v1.5 (384-dim)
LLM Claude (Anthropic API)
UI Framework Streamlit with NVIDIA dark theme
API Framework FastAPI
Export Formats Markdown, JSON, PDF, FHIR R4
Hardware Target NVIDIA DGX Spark

Agent Architecture Pattern

Every agent follows a consistent architecture:

User Query / Patient Data
    |
    v
[Domain-Specific Clinical Engines]
(deterministic scoring, classification, risk stratification)
    |
    v
[Multi-Collection RAG Search]
(BGE-small embedding → parallel Milvus search across N collections)
    |
    v
[Knowledge Graph Augmentation]
(domain entities, aliases, expansion maps)
    |
    v
[Claude LLM Synthesis]
(grounded response with citations)
    |
    v
[Clinical Output]
(Markdown | JSON | PDF | FHIR R4)

Key Numbers

Metric Value
Total agents 11
Total domain-specific collections 142
Shared genomic vectors 3.5M
Total test cases 6,000+
Therapeutic areas covered 13
Genes covered 201
Druggable targets 171
Query response time <5 seconds (vector search)

Getting Started

Clinical Decision Support Disclaimer

The Precision Intelligence Network agents are clinical decision support research tools. They are not FDA-cleared and are not intended as standalone diagnostic devices. All recommendations should be reviewed by qualified healthcare professionals. Apache 2.0 License.