A four-capstone portfolio progressing from optimization and backpropagation through transformer training to an auditable, MCP-based agentic incident-command system.
| Week | Capstone | Key concepts | Access |
|---|---|---|---|
| 1 | Gradient Descent Optimization | GD/SGD dynamics, learning-rate schedules, stochastic noise, and convergence diagnostics. |  |
| 2 | Backpropagation | Manual backpropagation, autograd validation, two-layer neural networks, and XOR classification. | |
| 3 | Tiny Transformer | BPE tokenization, causal self-attention, transformer training, and FOMC text generation. | |
| 4 | Agentic Incident Command | MCP client/server workflow, OPAL agent loop, guardrails, structured telemetry, and offline replay. |  |
the-ai-engineer/
├── capstones/
│ ├── week01_gd_optimization/ # GD/SGD notebook with inline figures
│ ├── week02_backprop/ # Manual → autograd → nn.Module
│ ├── week03_transformers/ # Tiny transformer + BPE extension
│ └── week04_agentic_incident_command/
│ ├── 01_tool_harness/ # Warm-up: minimal MCP server/client
│ ├── 02_incident_command_agent/ # Primary Week 4 capstone
│ ├── artifacts/ # Telemetry JSONL + sample summary
│ └── README_week04_capstone.md
├── pytest.ini
├── README.md
└── requirements.txt
This repository uses a lightweight pip + venv workflow and targets Python 3.11.
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtNotes:
- Weeks 1–3 can be reviewed directly on GitHub and opened in Google Colab using the badges above.
- Week 4 runs locally as a multi-process MCP application: one process hosts the server, another runs the agent/client, and the resulting telemetry is written to disk for deterministic replay and audit.
All content in this repository is provided for educational and illustrative purposes only. No guarantees are made regarding correctness, performance, reliability, or suitability for any production environment.
© 2026 Francisco Salazar