Skip to content

Week 13: Capstone Project Kickoff

Phase 5Week 13AdvancedLecture: 2026-05-26

Ralphthon is the capstone project for this course. Working individually, you will design and implement an autonomous agentic system that solves a real software problem, based on the Ralph Loop methodology.

The name is borrowed from the industry term “Hackathon.” As the Ralph Loop methodology spread in early 2026, Ralphthon culture emerged — and this is your chance to experience it firsthand.

Format

Individual projects. Implementation begins Week 13, based on the Week 8 approved proposal.

Topic

Keep the topic from your Week 8 proposal. Scope adjustments after feedback are allowed.

Beyond the cliché “automatic code review,” here are 6 ideas that attack novel problems with agentic systems. Selected so you can connect this course’s core techniques (loops, multi-agent, MCP, knowledge graphs) to real-world problems.

  1. Paper Reproducer Agent — Given an arXiv PDF, build a dependency environment, run the experiments, and reproduce the headline results (tables/figures). On failure, produce a diagnostic report pinpointing where the divergence occurred (environment, data, hyperparameters, or algorithm). Directly attacks the research reproducibility problem.

  2. Personal Knowledge Graph Distiller — Merge scattered notes from Obsidian, web bookmarks, and note apps into a Graphify-based knowledge graph, identify isolated “lone nodes,” and have an LLM propose new connections. Periodically auto-generates a “forgotten ideas recap” note. Applies the Week 5 LLM Wiki pattern.

  3. Docs-Code Drift Detector — Cross-verify claims in README, docstrings, and API docs against the actual code behavior at the AST level. When inconsistencies are found — e.g., “the doc says parse_json(str) → dict but the code returns a list” — auto-generate a fix PR. A doc-code consistency agent.

  4. Issue Roundtable — Given a GitHub Issue, have 4 agents (Planner/Architect/Coder/QA) debate the solution in an open-discussion format. When they disagree, log why; once a consensus design emerges, implement, test, and open a PR. The key is that the decision process is preserved as an auditable conversation log.

  5. Socratic Homework Tutor — Evaluate student-submitted code against a rubric but refuse to hand over the answer — instead, use Socratic questioning to guide thinking. Dynamically adjust hint strength based on the student’s responses, then summarize the path the student took to reach the answer. A deeper application of Week 6’s Learning output style.

  6. Dependency Upgrade Context Agent (Dependabot++) — For npm/pip dependency upgrade proposals, synthesize changelog, breaking changes, actual usage patterns, and test coverage into a “risk score + stepwise migration plan + expected edit locations.” Goes beyond Dependabot’s naive version bumps with context-aware upgrades.

  1. Problem Definition — Clearly define the real problem to solve and the criteria for success

  2. Agent Role Design — How many agents? What are each agent’s roles and responsibilities?

  3. Data Flow Design — Define the artifact formats passed between agents

  4. Harness Design — Backpressure mechanisms, garbage collection policies

  5. Infrastructure Design — Which MIG slices? Which models?

  6. Evaluation Method — How will the system’s performance be measured?

Submission deadline: 2026-06-02 23:59

Submission path: capstone/projects/[student-id]/design.md

Requirements:

  1. Problem definition and solution approach
  2. Agent architecture diagram
  3. Technology stack selection and rationale
  4. Week-by-week implementation plan
  5. Definition of success metrics