A POC proposal is a single markdown/slide doc following this skeleton:

# POC Title

Progress:

• Step 1: Identify POC category and target user (radiologist, tech, PACS admin)
• Step 2: Write problem statement with quantifiable pain point
• Step 3: Draft end-to-end workflow (input → AI processing → output → user action)
• Step 4: Design architecture diagram (edge inference vs cloud inference split)
• Step 5: Specify tech stack per layer (DICOM handling, AI serving, frontend viewer, infra)
• Step 6: List dependencies (datasets, DICOM conformance, hardware, third-party APIs)
• Step 7: Break effort into phases with man-month estimates
• Step 8: Define success metrics and demo scope (what "done" looks like for POC, not production)

Step 1: Identify POC Category

Map the request to one of the four archetypes (or a hybrid):

Step 2: Problem Statement Template

Today, [user role] must [manual task] which takes [time/cost].
This causes [downstream impact: delayed diagnosis, missed critical findings,
radiologist burnout, revenue leakage].
[Company]'s platform already handles [existing capability, e.g., DICOM routing,
teleradiology streaming] but lacks [AI capability gap].

Step 3: Workflow Template

Describe as a numbered pipeline, always including trigger, processing, and human-in-the-loop step:

1. DICOM study received at [edge gateway / cloud ingest]
2. Pre-processing: de-identification, modality/series filtering
3. AI inference: [model type] runs on [edge/cloud]
4. Output generated: [mask/contour/score/label] overlaid or attached as DICOM-SR/SEG
5. Result surfaced in [viewer/worklist] with confidence score
6. Radiologist reviews, accepts/overrides → feedback loop stored for retraining

Step 4: Architecture Diagram

Use a layered diagram (describe in text/mermaid, since this is a proposal doc):

MERMAID
flowchart LR
    A[Modality / Scanner] --> B[Edge Gateway - DICOM Receiver]
    B --> C{Inference Location}
    C -->|Low latency, PHI-sensitive| D[Edge AI Inference Container]
    C -->|Heavy compute, batch| E[Cloud AI Inference Service]
    D --> F[Web DICOM Viewer]
    E --> F
    F --> G[Radiologist Workstation / Browser]
    D --> H[Cloud Sync - Anonymized Metadata + Model Feedback]
    E --> H
    H --> I[Model Retraining Pipeline]

Always show: acquisition → edge → cloud boundary → viewer → feedback loop. Mark PHI boundary explicitly.

Step 5: Tech Stack Template

Organize by layer, give a recommended default plus alternative:

Step 6: Dependencies & Risks

Always cover these categories:

• Data: annotated dataset availability, IRB/data-use agreements, class imbalance for rare abnormalities
• Compliance: HIPAA/GDPR, de-identification pipeline, on-prem PHI restrictions for edge
• Hardware: GPU availability at edge sites, network bandwidth for cloud fallback
• Integration: existing PACS/RIS compatibility, DICOM conformance statement access
• Model: FDA/regulatory pathway if moving beyond POC (mention as future risk, not POC scope)

Step 7: Effort Breakdown Template (Man-Months)

Break into 4 phases; scale numbers to POC size (typically 2–4 months total, 1–3 engineers):

Always state team composition: e.g., "1 ML engineer, 1 full-stack/DICOM engineer, 0.5 clinical SME."

Step 8: Success Metrics

Give 3-5 measurable, POC-appropriate metrics (not production SLAs):

• Model accuracy/Dice score on held-out test set (e.g., Dice ≥ 0.85 for segmentation)
• Inference latency at edge (e.g., < 2s per series)
• Radiologist review time reduction (qualitative in POC, quantify in pilot)
• Successful DICOM round-trip with SEG/SR object generation

Example 1: DICOM Real-Time Web Viewer + Abnormality Detection (given example)

Input: "AI POC for real-time web viewer with segmentation mask/contour for abnormality detection, edge + cloud deployment."

Output:

# POC 1: Real-Time Interactive Reading with AI-Assisted Abnormality Detection

1. Study streamed to web viewer during live consult session
2. Edge gateway runs lightweight segmentation model (e.g., nnU-Net exported to ONNX) on incoming series in near real-time
3. Contour/mask overlay generated and rendered as DICOM-SEG object
4. Overlay streamed to viewer alongside live image, with confidence score badge
5. Radiologist toggles AI overlay on/off during consult, can annotate agreement/disagreement
6. Session feedback logged to cloud for model improvement

1. Study lands in cloud ingest queue
2. Triage classification model scores study (critical/urgent/routine) within seconds of upload
3. Worklist re-ranks studies by urgency score
4. Radiologist sees flagged critical studies at top with reason tag
5. Ground truth (final report) fed back to retrain triage model monthly