0.5 Capstone Project Thread: Course Knowledge Assistant

If you do not have your own project yet, default to a course knowledge assistant. It is not extra homework. It is one portfolio thread that grows through the course: each chapter adds one layer of capability until you have an AI application that can be explained, rerun, evaluated, and deployed.

Final Shape

By the end, this project should be able to:

• read course notes, PDFs, web excerpts, or your own learning records;
• clean data while preserving source, time, fields, and quality notes;
• answer with Prompt, RAG, or Agent workflows while keeping retrieval and tool traces;
• keep fixed eval questions, failure samples, cost/latency notes, and safety boundaries;
• optionally connect images, OCR, multimodal assets, or a local open-source model runtime;
• let a reviewer rerun the core path from the README.

Directory Template

capstone-course-assistant/
  README.md
  data/
    raw/
    processed/
  notebooks/
  src/
    cli.py
    data_pipeline.py
    evals.py
    rag.py
    agent_tools.py
  reports/
    evidence_log.md
    failure_cases.md
    eval_results.csv
    runtime_notes.md

On day one, only create the folder and README. The other files should appear naturally as the chapters add capability.

Portfolio Submission Template

Use the same final package format after every major stage. This keeps the project reviewable instead of becoming a pile of demos.

README.md                  what it does, how to run, what is not supported
run.sh or commands.md       exact rerun path
data_note.md                source, fields, cleaning rules, privacy notes
eval_cases.csv              fixed questions or inputs used for comparison
failure_cases.md            at least one honest failure and suspected cause
screenshots/ or outputs/    visible result, chart, trace, or API response
release_note.md             what changed this chapter and what to test next

Minimum version: README, one run command, one output, and one failure note. Strong portfolio version: fixed eval set, before/after comparison, cost or latency note, safety boundary, and a short demo script.

Growth By Chapter

Chapters 1-3: reproducible workbench Keep environment commands, Git commits, a Python CLI, sample data, cleaning rules, charts, and data quality notes.

Chapters 4-6: model evidence Use a small classification, regression, or representation experiment to practice baselines, metrics, failure samples, and training diagnosis. The goal is not a high score; the goal is evidence-based model judgment.

Chapter 7: LLM behavior control Fix 5-10 questions, then compare prompts, structured outputs, token/context limits, and failure samples. Optionally run mini GPT-2 to understand training and generation.

Chapter 8: RAG grounded answers Chunk course material, add metadata, retrieve evidence, and generate cited answers. Save top-k chunks before reading the final answer.

Chapter 9: Agent tool loop Expose only a few safe tools, such as reading files, listing folders, or generating reports. Keep tool schemas, traces, safety blocks, and rollback notes.

Chapters 10-12: product-specific extensions Use Chapter 10 for images or OCR, Chapter 11 for labels, extraction, or summaries, and Chapter 12 for PDF, image, audio, video, or creative-package workflows.

Chapter 13: open-source model runtime Start with a small model to run local inference, evaluation, and an OpenAI-style API. With GPU access, try vLLM or SGLang. Keep the model license, environment report, first run, eval table, and stop procedure.

Change One Thing Per Chapter

At the end of each chapter, answer four questions:

• What new capability did the project gain?
• What command reruns it?
• What evidence proves it works?
• What failure sample keeps the claim honest?

If you cannot answer, add evidence before adding features.

Evidence to Keep

Keep this page’s proof of learning as a project-thread evidence card:

Project Name

course knowledge assistant or your own replacement project

Chapter Growth Rule

add one capability per chapter, not a pile of demos

Rerun Path

README command, script, notebook cell, or service endpoint

Review Bundle

data note, eval cases, trace, failure note, and release note

Expected Output

one project thread that grows from setup to RAG, Agent, and runtime evidence

Minimum Pass Standard

After the main route, this project should include:

• a runnable README;
• a small dataset or document set;
• fixed evaluation questions;
• a Prompt/RAG/Agent trace;
• failure cases and an improvement plan;
• a note explaining when to use a cloud API and when to use an open-source model runtime.

The goal is not the largest system. The goal is a system that makes another person believe you understand the AI engineering loop.

Pass Check

You pass this planning page when your project thread has one rerun command, one evaluation artifact, and one known failure case.