Skip to content

Clinical Trial Intelligence Agent -- Demo Guide

Version: 2.0.0 Date: March 22, 2026 Author: Adam Jones Platform: NVIDIA DGX Spark -- HCLS AI Factory

Table of Contents

  1. Pre-Demo Checklist
  2. Starting the System
  3. UI Walkthrough: 5 Tabs
  4. Demo Scenario 1: Oncology Protocol Design
  5. Demo Scenario 2: Patient-Trial Matching
  6. Demo Scenario 3: Eligibility Optimization
  7. Demo Scenario 4: Competitive Intelligence
  8. Demo Scenario 5: Adaptive Design Selection
  9. API Demo Queries
  10. Talking Points
  11. FAQ
  12. Recovery Procedures

1. Pre-Demo Checklist

Run through this checklist before every demo:

# Check Command Expected
1 Tests pass python -m pytest tests/ -q 769 passed in 0.47s
2 API starts curl localhost:8538/health {"status": "ok"}
3 UI loads Open http://localhost:8128 NVIDIA-themed 5-tab UI
4 Collections exist curl localhost:8538/collections 14 collections listed
5 Knowledge loaded curl localhost:8538/v1/trial/knowledge-version version 2.0.0
6 Workflows available curl localhost:8538/workflows 10+ workflow types
7 Milvus running Check port 19530 Connected status in health
8 Browser configured Close unrelated tabs Clean browser window
9 Terminal ready Two terminals open API + backup commands
10 API key set Check .env ANTHROPIC_API_KEY present

2. Starting the System

Quick Start (Integrated)

cd /home/adam/projects/hcls-ai-factory
./start-factory.sh

Quick Start (Standalone)

cd /home/adam/projects/hcls-ai-factory/ai_agent_adds/clinical_trial_intelligence_agent

# Terminal 1: API
uvicorn api.main:app --host 0.0.0.0 --port 8538

# Terminal 2: UI
streamlit run app/trial_ui.py --server.port 8128

Verify Both Services

curl -s localhost:8538/health | python -m json.tool
# Open http://localhost:8128 in browser

3. UI Walkthrough: 5 Tabs

Tab 1: Trial Intelligence

Purpose: RAG-powered Q&A across all 14 collections with workflow-aware routing.

What to show: - Type a clinical trial question in the text input - System detects the appropriate workflow automatically - Results show: answer, citations with collection source and relevance scores, guideline references, confidence score - Point out the NVIDIA dark theme and responsive layout

Sample queries to type: - "What are the standard Phase 3 endpoints for NSCLC immunotherapy trials?" - "Compare adaptive design options for a rare disease Phase 2/3 study" - "What regulatory pathways are available for breakthrough therapy designation?"

Tab 2: Patient Matching

Purpose: Match a patient profile against clinical trial eligibility criteria.

What to show: - Enter patient demographics (age, sex, diagnosis) - Add biomarkers (e.g., PD-L1 TPS 80%, EGFR L858R) - Add genomic variants and current medications - Submit to see per-criterion match scores - Highlight the overall match score and confidence level - Show nearby trial sites if geographic location is provided

Tab 3: Protocol Optimizer

Purpose: Protocol complexity scoring and design recommendations.

What to show: - Enter trial parameters: indication, phase, procedure count, visit count, endpoint count, eligibility criteria count - Submit to see complexity score and percentile rank - Point out comparisons to Tufts CSDD industry benchmarks - Show endpoint recommendations by indication - Demonstrate historical success rate lookup

Tab 4: Competitive Landscape

Purpose: Competitive threat assessment and landscape visualization.

What to show: - Enter an indication and mechanism of action - See competitor trial listing with threat scores - Explain the 4-factor threat model (phase, enrollment, sponsor, differentiation) - Highlight threat classification (critical/high/moderate/low/minimal) - Show enrollment progress comparison

Tab 5: Dashboard

Purpose: System health and operational metrics.

What to show: - Collection health indicators (14 collections with record counts) - Query volume metrics - Workflow execution breakdown - Cross-agent integration status - Knowledge base version and last update

4. Demo Scenario 1: Oncology Protocol Design

Story: "A pharma company wants to design a Phase 3 trial for a novel PD-L1/VEGF bispecific antibody in first-line NSCLC. Let's use the agent to generate an evidence-based protocol blueprint."

Step 1: Ask the Question

Go to Tab 1 (Trial Intelligence) and type:

"Design a Phase 3 protocol for a PD-L1/VEGF bispecific antibody in first-line metastatic NSCLC with PD-L1 TPS >= 50%"

Expected Output

The system should return:

  • Workflow detected: Protocol Design
  • Recommended primary endpoint: Progression-Free Survival (PFS) or Overall Survival (OS)
  • Recommended secondary endpoints: ORR, DoR, OS (if PFS is primary), quality of life
  • Comparator recommendation: Pembrolizumab monotherapy (based on KEYNOTE-024)
  • Sample size estimate: 300-600 patients based on expected HR and event rate
  • Eligibility framework: PD-L1 TPS >= 50%, ECOG 0-1, no prior systemic therapy
  • Adaptive design suggestion: Group sequential with interim analysis for OS
  • Regulatory pathway: Potential for Breakthrough Therapy Designation; RTOR eligible
  • Landmark trial references: KEYNOTE-024 (pembrolizumab), IMpower110 (atezolizumab)
  • Historical success rate: ~36% for Phase 3 oncology

Talking Points

  • "Notice how the system referenced KEYNOTE-024 as the benchmark trial -- that's the current standard of care for this indication"
  • "The PFS and OS dual primary endpoint strategy is exactly what FDA expects for first-line NSCLC"
  • "The 36% historical success rate is calibrated from BIO/QLS Advisors data across thousands of oncology trials"

5. Demo Scenario 2: Patient-Trial Matching

Story: "A 58-year-old male with metastatic NSCLC, PD-L1 TPS 80%, EGFR wild-type, no prior systemic therapy, and ECOG 1 wants to find matching clinical trials."

Step 1: Enter Patient Profile

Go to Tab 2 (Patient Matching) and enter:

Field Value
Age 58
Sex Male
Diagnosis Metastatic non-small cell lung cancer
Biomarkers PD-L1 TPS 80%, EGFR wild-type, ALK negative
Medications None (treatment-naive)
Genomic variants KRAS G12C
Comorbidities Hypertension (controlled)
Location Boston, MA

Expected Output

  • Multiple trial matches with per-criterion scoring
  • High match score for first-line IO trials (PD-L1 >= 50%)
  • KRAS G12C targeted therapy trials flagged
  • Controlled hypertension should pass cardiac exclusion criteria
  • Nearby Boston-area trial sites listed

Talking Points

  • "Each eligibility criterion gets its own confidence score -- the system doesn't just say yes or no"
  • "Notice the KRAS G12C variant triggered an additional match for targeted therapy trials"
  • "The cross-agent trigger sent this to the oncology agent for molecular match confirmation"

6. Demo Scenario 3: Eligibility Optimization

Story: "A trial protocol has 28 eligibility criteria and is struggling with enrollment. The medical team wants to know which criteria could be broadened without compromising safety."

Step 1: Submit Criteria for Analysis

Go to Tab 1 (Trial Intelligence) and type:

"Analyze these eligibility criteria for enrollment impact: ECOG 0-1, no prior immunotherapy, no CNS metastases, hemoglobin >= 10 g/dL, no autoimmune disease, creatinine clearance >= 30 mL/min, no prior organ transplant, ejection fraction >= 50%, pregnancy or lactation excluded"

Expected Output

The system should return criteria ranked by population impact:

Criterion Population Impact Recommendation
Pregnancy or lactation 50% RETAIN with monitoring
ECOG 0-1 25% REVIEW: consider ECOG 0-2
No prior immunotherapy 20% BROADEN: weak justification for some lines
No CNS metastases 15% REVIEW: stable CNS may be includable
Ejection fraction >= 50% 8% RETAIN: cardiac safety justified
Creatinine clearance >= 30 15% REVIEW: could lower threshold
No autoimmune disease 8% REVIEW: mild autoimmune may be safe

Talking Points

  • "The system identified that 'no prior immunotherapy' excludes 20% of the population but has weak scientific justification for this particular trial"
  • "Each criterion's population impact is calibrated against real-world data -- not guesswork"
  • "RETAIN recommendations have strong scientific backing; BROADEN recommendations have high impact with weak justification"

7. Demo Scenario 4: Competitive Intelligence

Story: "The sponsor wants to understand the competitive landscape for GLP-1 receptor agonists in obesity before starting their Phase 3 program."

Step 1: Query the Landscape

Go to Tab 4 (Competitive Landscape) or type in Tab 1:

"What is the competitive landscape for GLP-1 agonists in obesity Phase 3 trials? Include semaglutide, tirzepatide, and survodutide."

Expected Output

  • Competitor profiles for SURMOUNT-1 (tirzepatide), SELECT (semaglutide), and others
  • Threat scores with classification:
  • Tirzepatide/Zepbound: Critical (0.85+) -- Phase 3 completed, 22.5% weight loss
  • Semaglutide/Wegovy: Critical (0.90+) -- Approved, MACE benefit proven
  • Survodutide: High (0.65) -- Phase 3 ongoing
  • Enrollment progress comparison
  • Differentiation analysis (mechanism similarity)

Talking Points

  • "The threat model considers four factors: phase, enrollment, sponsor resources, and mechanism similarity"
  • "Semaglutide scores as critical because it's already approved AND has cardiovascular outcome data"
  • "This analysis helps sponsors decide whether to differentiate on efficacy, safety, convenience, or a new endpoint"

8. Demo Scenario 5: Adaptive Design Selection

Story: "A biotech company is developing a novel gene therapy for a rare disease (hemophilia A) and wants to know which adaptive design is best for their Phase 2/3 program."

Step 1: Ask About Adaptive Designs

Type in Tab 1:

"What adaptive trial design should I use for a Phase 2/3 seamless gene therapy trial in hemophilia A? Consider the small patient population and uncertain dose-response."

Expected Output

  • Recommended design: Seamless Phase 2/3 with sample size re-estimation
  • Alternative designs: Response-adaptive randomization, biomarker-adaptive enrichment
  • Rationale: Small population (rare disease), uncertain effect size, long follow-up needed
  • Regulatory guidance: FDA Guidance on Adaptive Designs (2019), RMAT designation eligible
  • Precedent trials: SPRINT_SMA (gene therapy adaptive design), Zolgensma experience
  • Statistical considerations: Bayesian framework, external historical controls, LTFU requirements
  • Success rate estimate: ~55% for Phase 3 rare disease (higher than average)

Talking Points

  • "The system recommended seamless Phase 2/3 because it eliminates the inter-trial gap -- critical for rare diseases where every patient matters"
  • "FDA RMAT designation makes this eligible for accelerated timelines"
  • "The 55% Phase 3 success rate for rare diseases is significantly higher than the overall average of ~50%, reflecting smaller trials and often dramatic clinical effects"

9. API Demo Queries

For technical audiences, demonstrate the REST API directly:

9.1 RAG Query

curl -X POST http://localhost:8538/v1/trial/query \
    -H "Content-Type: application/json" \
    -d '{
        "question": "What biomarker strategies are used in NSCLC immunotherapy trials?",
        "workflow_type": "protocol_design",
        "top_k": 5
    }'

9.2 Patient Matching

curl -X POST http://localhost:8538/v1/trial/match \
    -H "Content-Type: application/json" \
    -d '{
        "patient": {
            "age": 65,
            "sex": "female",
            "diagnosis": "HER2-low metastatic breast cancer",
            "biomarkers": ["HER2 IHC 1+", "HR-positive", "BRCA wild-type"]
        }
    }'

9.3 Safety Signal Detection

curl -X POST http://localhost:8538/v1/trial/safety/signal \
    -H "Content-Type: application/json" \
    -d '{
        "events": [
            {"event_type": "hepatotoxicity", "severity": "grade_3", "frequency": 0.08},
            {"event_type": "neutropenia", "severity": "grade_4", "frequency": 0.12}
        ],
        "trial_id": "NCT00000001"
    }'

9.4 Knowledge Version

curl -s http://localhost:8538/v1/trial/knowledge-version | python -m json.tool

9.5 Therapeutic Areas

curl -s http://localhost:8538/v1/trial/therapeutic-areas | python -m json.tool

10. Talking Points

For Executive Audiences

  • "This agent reduces the time to generate an evidence-based protocol blueprint from weeks to seconds"
  • "40 landmark trials from KEYNOTE-024 to RECOVERY to CLARITY-AD serve as real-world design precedents"
  • "The system covers 13 therapeutic areas, from oncology to gene therapy, with calibrated success rates"
  • "Running on DGX Spark, this integrates with genomics, drug discovery, and four peer AI agents"

For Clinical Operations Audiences

  • "Patient-trial matching evaluates every eligibility criterion with confidence scoring"
  • "Eligibility optimization identifies criteria that exclude patients without scientific justification"
  • "Site selection considers enrollment history, diversity index, and screen failure rates"
  • "Enrollment predictions use prevalence, competition, and phase-specific factors"

For Regulatory Affairs Audiences

  • "The system references ICH E6(R3), E9(R1), and E8(R1) guidelines"
  • "9 regulatory agencies covered with approval pathways and expedited programs"
  • "Adaptive design recommendations include specific FDA guidance citations"
  • "Every recommendation includes evidence level classification (A1 through E)"

For Data Science Audiences

  • "14 Milvus vector collections with IVF_FLAT indexing and COSINE similarity"
  • "384-dimensional BGE-small-en-v1.5 embeddings for clinical trial text"
  • "Workflow-specific collection weight boosting for domain-relevant retrieval"
  • "Calibrated confidence model: evidence base (0.3) + raw confidence (0.3) + documents (0.2) + agreement (0.2)"
  • "769 tests in 0.47 seconds -- 100% pass rate"

11. FAQ

Q: Does this replace clinical trial design teams? A: No. This is a decision support tool that accelerates evidence retrieval and provides calibrated recommendations. All outputs should be reviewed by qualified clinical, regulatory, and statistical professionals.

Q: How current is the trial data? A: The knowledge base contains 40 curated landmark trials and comprehensive reference data. Real-time ClinicalTrials.gov integration is available through the ingest pipeline on a configurable schedule (default 24 hours).

Q: What happens if Milvus is down? A: The system degrades gracefully. All 10 workflows, 5 decision support engines, and the knowledge base continue to function. Only vector search results are unavailable.

Q: What happens without an LLM API key? A: The system operates in search-only mode. Vector searches, workflow execution, and decision support engines all function normally. Natural language synthesis is unavailable.

Q: Can this be used for regulatory submissions? A: The system generates draft documents and evidence summaries. All regulatory submissions require expert review, validation, and formatting per agency-specific requirements.

Q: What therapeutic areas are supported? A: All 13 major therapeutic areas: oncology, cardiovascular, neuroscience, immunology, infectious disease, rare diseases, metabolic/endocrinology, respiratory, hematology, gastroenterology, dermatology, ophthalmology, and gene/cell therapy.

12. Recovery Procedures

If the API Crashes

# Restart API
uvicorn api.main:app --host 0.0.0.0 --port 8538

If the UI Freezes

# Restart Streamlit (it auto-reconnects)
streamlit run app/trial_ui.py --server.port 8128

If Milvus Is Unreachable

# Check Milvus status
docker ps | grep milvus

# Restart if needed
docker restart milvus-standalone

# Verify
curl localhost:8538/health

If Searches Return Empty

# Re-seed knowledge base
python scripts/seed_knowledge.py

# Re-create collections if needed
python scripts/setup_collections.py
python scripts/seed_knowledge.py

Clinical Trial Intelligence Agent v2.0.0 -- Demo Guide -- March 2026