WaferGuard ML - Project Blueprint

Team: Milos, Hernan, Rolando, Oliver Duration: 10-12 weeks (5-6 sprints) Sprint Length: 2 weeks

---

System Architecture

flowchart TB
    subgraph Data["Data Ingestion Layer"]
        SECOM[(SECOM Sensor Data)]
        WM811K[(WM811K Wafer Images)]
        VAL[Data Validation]
    end

    subgraph ML["Machine Learning Pipeline"]
        PRE[Preprocessing]
        FE[Feature Engineering]
        CNN[CNN Model]
        AE[Autoencoder]
        ENS[Ensemble Model]
    end

    subgraph INF["Inference & Alert System"]
        API[FastAPI Endpoint]
        CONF[Confidence Scoring]
        ALERT[Alert Generator]
    end

    subgraph VIZ["Visualization Dashboard"]
        STREAM[Streamlit Dashboard]
        KPI[KPI Metrics]
        GCAM[GradCAM Views]
        HIST[Historical Trends]
    end

    subgraph MES["MES Integration"]
        SIM[MES Simulator]
        HOOK[Integration Hooks]
    end

    SECOM --> VAL
    WM811K --> VAL
    VAL --> PRE
    PRE --> FE
    FE --> CNN
    FE --> AE
    CNN --> ENS
    AE --> ENS
    ENS --> API
    API --> CONF
    CONF --> ALERT
    API --> STREAM
    ALERT --> STREAM
    STREAM --> KPI
    STREAM --> GCAM
    STREAM --> HIST
    API --> HOOK
    SIM --> HOOK

---

Data Flow

flowchart LR
    A[Raw Data] --> B[Preprocessing]
    B --> C[Feature Engineering]
    C --> D[Model Training]
    D --> E[Validation]
    E --> F[Deployment]
    F --> G[Real-time Inference]
    G --> H[Dashboard]
    H --> I[Alerts]

    style A fill:#e1f5fe
    style D fill:#fff3e0
    style H fill:#e8f5e9
    style I fill:#ffebee

---

Epics Overview

Recommended Epic Structure

Epic ID	Name	Description	Phase Alignment
KAN-1	Data Pipeline & Foundation	Environment setup, data acquisition, EDA, preprocessing	Phase 1-2
KAN-2	Anomaly Detection (ML) Model	CNN + Autoencoder development, training, validation	Phase 3
KAN-3	MES Integration & API	FastAPI inference endpoints, MES simulation, alert system	Phase 4
KAN-4	Analytics + Dashboard	Streamlit dashboard, visualizations, KPIs, GradCAM	Phase 4
KAN-5	Governance & Audit-ability	Model versioning (MLflow), documentation, evaluation reports	Phase 5

Why This Structure?

Your original epics were solid. I recommend: - Split KAN-1 into Data Pipeline (new KAN-1) and ML Model (keep as KAN-2) - Keep KAN-2 (MES Integration) as KAN-3 - Keep KAN-3 (Analytics+Dashboard) as KAN-4 - Keep KAN-4 (Governance) as KAN-5

This gives you a clear dependency chain: Data → Model → Integration → Dashboard → Documentation

Epic Dependency Flow

flowchart LR
    KAN1[KAN-1<br/>Data Pipeline]
    KAN2[KAN-2<br/>ML Model]
    KAN3[KAN-3<br/>MES Integration]
    KAN4[KAN-4<br/>Dashboard]
    KAN5[KAN-5<br/>Governance]

    KAN1 --> KAN2
    KAN2 --> KAN3
    KAN2 --> KAN4
    KAN3 --> KAN4
    KAN4 --> KAN5
    KAN3 --> KAN5

    style KAN1 fill:#bbdefb,color:#000000
    style KAN2 fill:#c8e6c9,color:#000000
    style KAN3 fill:#fff9c4,color:#000000
    style KAN4 fill:#ffccbc,color:#000000
    style KAN5 fill:#e1bee7,color:#000000

---

Sprint Timeline

gantt
    title WaferGuard ML - Sprint Timeline
    dateFormat  YYYY-MM-DD
    section KAN-1 Data Pipeline
        Sprint 1 - Foundation     :s1, 2026-01-27, 14d
        Sprint 2 - Data Pipeline  :s2, after s1, 14d
    section KAN-2 ML Model
        Sprint 3 - Model Dev      :s3, after s2, 14d
        Sprint 4 - Refinement     :s4, after s3, 14d
    section KAN-3 & KAN-4
        Sprint 4 - API            :s4b, after s3, 14d
        Sprint 5 - Dashboard      :s5, after s4, 14d
    section KAN-5 Governance
        Sprint 6 - Documentation  :s6, after s5, 14d

---

Sprint Breakdown

Sprint 1: Foundation (Week 1-2)

Epic: KAN-1 (Data Pipeline & Foundation) - 80% Done

Task ID	Task	Owner	Status
KAN-8	Set up Python Environment	Oliver	DONE ✅
KAN-9	Initialize Git Repo With Branching Strategy + Rules	Oliver	DONE ✅
KAN-10	Download Secom Dataset + Verify Integrity	Oliver	DONE ✅
KAN-11	Download WM811 Image Analysis + Verify	Oliver	DONE ✅
KAN-12	Run EDA on Secom Sensor Data	Rolando	IN REVIEW 🔄
KAN-13	Run EDA on WM811K Image analysis	Oliver	DONE ✅
KAN-14	Data Quality Assessment Report	Unassigned	TO DO 📋
KAN-15	Set Up Jira Board with epics/stories	Rolando	DONE ✅
KAN-46	Research about how the pixels are generated from and actual image	Milos	DONE ✅
KAN-47	Create slides for initial images EDA findings	Oliver	DONE ✅

Sprint 1 Goal: Development environment ready, datasets loaded, initial EDA complete

---

Sprint 2: Data Pipeline (Week 3-4)

Epic: KAN-1 (Data Pipeline & Foundation)

Task ID	Task	Owner	Story Points	Status
KAN-1-010	SECOM preprocessing pipeline (missing values, outliers)	TBD	5	TODO
KAN-1-011	Feature engineering for sensor data	TBD	5	TODO
KAN-1-012	WM811K image preprocessing (resize, normalize)	TBD	3	TODO
KAN-1-013	Image augmentation strategy (rotation, flip)	TBD	3	TODO
KAN-1-014	Train/validation/test split implementation	TBD	2	TODO
KAN-1-015	Data loader classes (PyTorch/TensorFlow)	TBD	5	TODO
KAN-1-016	Handle class imbalance (SMOTE/class weights)	TBD	3	TODO

Sprint 2 Goal: Clean, processed datasets ready for model training

---

Sprint 3: ML Model Development (Week 5-6)

Epic: KAN-2 (Anomaly Detection Model)

Task ID	Task	Owner	Story Points	Status
KAN-2-001	CNN architecture design for wafer classification	TBD	5	TODO
KAN-2-002	Implement CNN model (PyTorch/TensorFlow)	TBD	5	TODO
KAN-2-003	Autoencoder architecture for sensor anomaly	TBD	5	TODO
KAN-2-004	Implement Autoencoder model	TBD	5	TODO
KAN-2-005	Training pipeline with early stopping	TBD	3	TODO
KAN-2-006	Hyperparameter tuning setup	TBD	3	TODO
KAN-2-007	Baseline model benchmarking	TBD	3	TODO

Sprint 3 Goal: Working CNN and Autoencoder models with baseline metrics

---

Sprint 4: Model Refinement + API (Week 7-8)

Epics: KAN-2 (Model), KAN-3 (MES Integration)

Task ID	Task	Owner	Story Points	Status
KAN-2-008	Model validation and cross-validation	TBD	3	TODO
KAN-2-009	Ensemble model (CNN + Autoencoder)	TBD	5	TODO
KAN-2-010	GradCAM implementation for explainability	TBD	5	TODO
KAN-3-001	FastAPI inference endpoint design	TBD	3	TODO
KAN-3-002	Implement /predict endpoint	TBD	5	TODO
KAN-3-003	Confidence scoring and threshold logic	TBD	3	TODO
KAN-3-004	MES simulation mock data generator	TBD	3	TODO
KAN-3-005	Alert generation logic	TBD	3	TODO

Sprint 4 Goal: Optimized models, working inference API

---

Sprint 5: Dashboard Development (Week 9-10)

Epic: KAN-4 (Analytics + Dashboard)

Task ID	Task	Owner	Story Points	Status
KAN-4-001	Dashboard wireframes in Figma	TBD	3	TODO
KAN-4-002	Streamlit app skeleton	TBD	2	TODO
KAN-4-003	Real-time anomaly detection view	TBD	5	TODO
KAN-4-004	Historical trend analysis charts	TBD	5	TODO
KAN-4-005	Model confidence visualization	TBD	3	TODO
KAN-4-006	GradCAM attention map display	TBD	5	TODO
KAN-4-007	Production KPIs (OEE, defect rate)	TBD	3	TODO
KAN-4-008	Alert notification panel	TBD	3	TODO
KAN-4-009	Interactive filtering and drill-down	TBD	5	TODO

Sprint 5 Goal: Functional dashboard with all visualizations

---

Sprint 6: Evaluation & Documentation (Week 11-12)

Epic: KAN-5 (Governance & Audit-ability)

Task ID	Task	Owner	Story Points	Status
KAN-5-001	Comprehensive model evaluation report	TBD	5	TODO
KAN-5-002	Comparison vs baseline (SPC, manual)	TBD	5	TODO
KAN-5-003	Model versioning documentation	TBD	3	TODO
KAN-5-004	API documentation (OpenAPI/Swagger)	TBD	3	TODO
KAN-5-005	User guide for dashboard	TBD	3	TODO
KAN-5-006	Technical architecture documentation	TBD	3	TODO
KAN-5-007	Final presentation slides	TBD	5	TODO
KAN-5-008	Live demo preparation	TBD	3	TODO
KAN-5-009	Code cleanup and refactoring	TBD	3	TODO

Sprint 6 Goal: Complete documentation, presentation ready

---

Work Stream Assignments

Work Stream	Specialist	Assistant	Primary Epics
Data Engineering & ML Pipeline	TBD	TBD	KAN-1
Model Development & Training	TBD	TBD	KAN-2
Dashboard & Integration	TBD	TBD	KAN-3, KAN-4
Evaluation & Documentation	TBD	TBD	KAN-5

Team Structure

flowchart TB
    subgraph Team["WaferGuard ML Team"]
        subgraph WS1["Data Engineering"]
            DE1[Specialist: TBD]
            DE2[Assistant: TBD]
        end
        subgraph WS2["Model Development"]
            MD1[Specialist: TBD]
            MD2[Assistant: TBD]
        end
        subgraph WS3["Dashboard & Integration"]
            DI1[Specialist: TBD]
            DI2[Assistant: TBD]
        end
        subgraph WS4["Evaluation & Docs"]
            ED1[Specialist: TBD]
            ED2[Assistant: TBD]
        end
    end

    WS1 -->|Data| WS2
    WS2 -->|Models| WS3
    WS3 -->|System| WS4

    style WS1 fill:#bbdefb
    style WS2 fill:#c8e6c9
    style WS3 fill:#fff9c4
    style WS4 fill:#e1bee7

---

ML Model Architecture

flowchart TB
    subgraph Input["Input Data"]
        IMG[Wafer Images<br/>WM811K]
        SEN[Sensor Data<br/>SECOM]
    end

    subgraph CNN["CNN Pipeline"]
        C1[Conv2D + ReLU]
        C2[MaxPool]
        C3[Conv2D + ReLU]
        C4[MaxPool]
        C5[Flatten]
        C6[Dense + Softmax]
    end

    subgraph AE["Autoencoder Pipeline"]
        E1[Encoder]
        E2[Latent Space]
        E3[Decoder]
        E4[Reconstruction Error]
    end

    subgraph Ensemble["Ensemble Decision"]
        COMB[Score Combination]
        THRESH[Threshold Logic]
        OUT[Final Prediction]
    end

    IMG --> C1 --> C2 --> C3 --> C4 --> C5 --> C6
    SEN --> E1 --> E2 --> E3 --> E4
    C6 --> COMB
    E4 --> COMB
    COMB --> THRESH --> OUT

    style CNN fill:#e3f2fd
    style AE fill:#f3e5f5
    style Ensemble fill:#e8f5e9

---

Technology Stack (Recommended)

Core:           Python 3.9+
ML Framework:   PyTorch (recommended) or TensorFlow/Keras
Data:           NumPy, Pandas, Scikit-learn, OpenCV
Dashboard:      Streamlit (faster dev) or Plotly Dash
API:            FastAPI
Tracking:       MLflow
Explainability: SHAP, GradCAM
Version Control: Git + GitHub
Project Mgmt:   Jira

Tech Stack Diagram

flowchart LR
    subgraph Frontend["Frontend"]
        ST[Streamlit]
        FIG[Figma]
    end

    subgraph Backend["Backend"]
        FA[FastAPI]
        ML[MLflow]
    end

    subgraph ML_Stack["ML Stack"]
        PT[PyTorch]
        SK[Scikit-learn]
        CV[OpenCV]
    end

    subgraph Data_Stack["Data"]
        PD[Pandas]
        NP[NumPy]
    end

    subgraph DevOps["DevOps"]
        GIT[Git/GitHub]
        JIRA[Jira]
    end

    Data_Stack --> ML_Stack --> Backend --> Frontend
    DevOps -.-> Backend
    DevOps -.-> ML_Stack

---

Key Datasets

Dataset	Description	Size	Source
SECOM	Sensor data (590 features)	~1,500 samples	UCI ML Repository
WM811K	Wafer map images	~38,000 images	Kaggle

Data Pipeline Overview

flowchart LR
    subgraph Sources["Data Sources"]
        S1[(SECOM<br/>UCI Repository)]
        S2[(WM811K<br/>Kaggle)]
    end

    subgraph Processing["Processing"]
        P1[Missing Value<br/>Imputation]
        P2[Outlier<br/>Detection]
        P3[Normalization]
        P4[Image Resize<br/>& Augment]
    end

    subgraph Output["Processed Data"]
        O1[Train Set<br/>70%]
        O2[Val Set<br/>15%]
        O3[Test Set<br/>15%]
    end

    S1 --> P1 --> P2 --> P3 --> O1 & O2 & O3
    S2 --> P4 --> O1 & O2 & O3

    style Sources fill:#e3f2fd
    style Processing fill:#fff8e1
    style Output fill:#e8f5e9

---

Success Criteria

Metric	Target
CNN Accuracy	>90%
False Positive Rate	<5%
Inference Latency	<500ms
Dashboard Refresh	<2s

Story Points by Epic

pie showData
    title Story Points Distribution
    "KAN-1 Data Pipeline" : 47
    "KAN-2 ML Model" : 42
    "KAN-3 MES Integration" : 17
    "KAN-4 Dashboard" : 34
    "KAN-5 Governance" : 33

---

Inference Flow

sequenceDiagram
    participant MES as MES System
    participant API as FastAPI
    participant Model as ML Model
    participant DB as MLflow
    participant Dash as Dashboard
    participant Eng as Process Engineer

    MES->>API: POST /predict (wafer data)
    API->>Model: Load model from registry
    Model->>DB: Get latest model version
    DB-->>Model: Return model weights
    Model->>Model: Run inference
    Model-->>API: Prediction + confidence
    API->>API: Apply threshold logic

    alt Anomaly Detected
        API->>Dash: Send alert
        Dash->>Eng: Display notification
        API-->>MES: Response (anomaly: true)
    else Normal
        API-->>MES: Response (anomaly: false)
    end

    API->>Dash: Update real-time view

---

Sprint Workflow

stateDiagram-v2
    [*] --> Planning: Sprint Start
    Planning --> InProgress: Tasks Assigned
    InProgress --> CodeReview: PR Created
    CodeReview --> Testing: Approved
    Testing --> Done: Tests Pass
    CodeReview --> InProgress: Changes Requested
    Testing --> InProgress: Tests Fail
    Done --> [*]: Sprint End

    note right of Planning: 2 hour session
    note right of InProgress: Daily standups
    note right of Done: Sprint review

---

Current Sprint Status

Active Sprint: Sprint 1 - Foundation Sprint Start: TBD Sprint End: TBD

Sprint 1 Progress

• Environment setup
• Data acquisition
• EDA complete
• Jira configured

---

Meeting Schedule

Meeting	Frequency	Duration	Day/Time
Sprint Planning	Bi-weekly	2 hours	Sprint Day 1
Daily Standup	Daily	15 min	TBD
Sprint Review	Bi-weekly	1.5 hours	Sprint Last Day
Sprint Retro	Bi-weekly	1 hour	Sprint Last Day

---

Quick Links

• Jira Board: TBD
• GitHub Repo: TBD
• Figma Designs: TBD
• MLflow Dashboard: TBD
• Google Drive: TBD

---

Last Updated: 2026-01-23