APPENDIX H: RESUME PORTFOLIO BUILDING CAPSTONE PROJECTS

HOW TO PRESENT IT IN A PORTFOLIO README outline Title and a one-sentence purpose that mentions context management. The problem in plain words: small desk, many books. A text diagram showing the flow from documents to notes to answer, with compaction in the middle. How to run the project with the mock model. How compaction works and what the limit is set to. How the restart works and how the test proves it. Honest limits, including that the mock extracts marked fact lines and the size count is only a rough guess. Next steps. What to screenshot or record Take a screenshot of output/notes.md next to output/answer.txt. Take a screenshot of the log lines where the compacted events appear. Record a short screen video of the run being stopped after six steps and then restarted, with the terminal showing that only four documents were read the second time. Metrics to report honestly Report the token limit you used and how many times compaction happened in a full run. Report how many documents were read in a fresh run compared with a restarted run, which should be six and four. Report that the numbers come from invented documents and a mock model, and note that your size count is an estimate. How to word a resume bullet truthfully Good example: "Built a context manager for a Python research agent that compacts old conversation at a token limit and persists findings to a notes file; tests show a restarted run skips already-finished documents." Another good example: "Designed a notes-based resume-from-failure mechanism for an agent loop, verified with automated tests on synthetic documents." Avoid saying "solved long-term memory for AI agents," because the project demonstrates one simple pattern on a tiny example. Interview talking points Explain why a model cannot simply be given everything, using the small desk analogy. Explain the trade-off in compaction, which is that every summary loses detail. Explain why you kept the facts in a file outside the conversation. Explain how you would choose a real token limit for a real model. Describe how you would detect that a summary dropped something important.