Bob The Builder

IBM Bob as a Senior Deep Learning Engineer

---

Overview

Bob The Builder is an AI agent that uses IBM Bob to automatically optimize deep learning models for CIFAR-100 image classification. Bob acts like a senior deep learning engineer — it experiments with hyperparameters, architecture changes, and regularization strategies, automatically committing improvements with Git and rolling back failures.

---

How It Works

1. Bob reads the training code (train.py)
2. Bob suggests a hyperparameter change
3. Training runs for 5–10 minutes
4. If accuracy improves → Bob commits with git commit
5. If accuracy drops → Bob rolls back with git reset --hard
6. Steps 2–5 repeat for 30 iterations
7. A Streamlit dashboard shows real-time progress

---

Optimizations Bob Performs

Category	Options Tested
Learning Rate	0.0001, 0.0005, 0.001, 0.005, 0.01
Batch Size	32, 64, 128
Optimizer	Adam, SGD, AdamW
Dropout Rate	0.0, 0.2, 0.3, 0.5
Batch Normalization	True, False
Weight Decay	0.0, 1e-5, 1e-4, 1e-3
Data Augmentation	True, False

Total possible combinations: 1,000+  |  Actual experiments run: 30

---

Results

Metric	Value
Baseline Accuracy	40.43%
Best Accuracy	44.67%
Total Improvement	+4.24%
Relative Gain	+10.5%
Winning Change	Batch Normalization (+3.91%)

Key Discoveries

Outcome	Change	Notes
✅ Success	Batch Normalization	+3.91% improvement
❌ Failed	Dropout	Made results worse
❌ Failed	Weight Decay	No improvement
❌ Failed	SGD Optimizer	Only 27.3% accuracy
❌ Failed	Higher Learning Rate	Made results worse
❌ Failed	Data Augmentation	No gain

Conclusion

Bob discovered that for CIFAR-100 with a CNN architecture, Batch Normalization alone provides the biggest improvement. Adding complexity — dropout, augmentation, different optimizers — makes results worse. Simplicity wins.

---

Tech Stack

• Python 3.9+
• PyTorch
• Streamlit
• Git
• IBM Bob

---

Project Structure

Bob-The-Builder/
├── train.py           # CNN model for CIFAR-100
├── optimizer.py       # Bob's control loop + Git automation
├── config.py          # Configuration (read-only)
├── objective.py       # Scoring logic for improvements
├── architecture.py    # Flexible CNN architecture
├── augmentation.py    # Data augmentation strategies
├── training.py        # Training loop with early stopping
├── dashboard.py       # Streamlit dashboard
├── requirements.txt   # Python dependencies
├── README.md
├── experiments.csv    # All 30 experiment logs
├── best_model.pth     # Best trained model (44.67%)
└── data/              # CIFAR-100 dataset

---

Setup & Usage

# 1. Clone the repository
git clone https://github.com/vivek41-glitch/Bob-The-Builder.git
cd Bob-The-Builder

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run baseline (optional)
python train.py

# 4. Run Bob's optimization
python optimizer.py

# 5. Launch dashboard
streamlit run dashboard.py

---

Expected Results

Phase	Accuracy
Baseline	40–45%
After Optimization	44–50%
Improvement	+4–5%

---

Team — Team Destroyer

Role	Name
Team Lead & ML Engineer	Vivek Dubey
Frontend & Dashboard	Areeba Abid
Model Architecture	Atikah Qaisar
Training Pipeline	Sam (SUDO)
Documentation & QA	Rashmika

---

Links

• GitHub: github.com/vivek41-glitch/Bob-The-Builder

---

License

MIT License — free to use, modify, and distribute.

---

Bob The Builder — Can he build it? Yes he can.