Pre-semester is a real thing
I got my acceptance to the CMU MS in AI Engineering, ECE track, and then I had four months of runway before the first day of class in August 2026. Nobody warned me that the pre-semester phase would have its own texture. You are not a student yet, but the program has already changed how you read papers. You are not looking for anything specific, but you are looking. The bookmarks folder swells. The unread arXiv tabs get to twelve, then twenty, then you start closing them without reading because the churn is faster than the throughput.
Three private repos on my GitHub map onto this stretch. Two of them are Slidev decks for a portfolio talk I have not given yet. The third is a JavaScript project that is really a research artifact: a directory of every person I could find who has already been through the ECE MS AIE program.
None of these are polished. That is the point of writing about them. Pre-semester work is embarrassing in the specific way that all preparation is: you are optimizing for a room whose furniture you have not seen.
The two Slidev repos
The two repos are cmuSlidev (created 2026-04-03 06:30 UTC) and SlidevCMU (created 2026-04-03 09:34 UTC). About three hours apart. Both are Vue-primary. Both have a single "Initial commit". Both have the same slides.md.
The second one is a rename. I made cmuSlidev, opened it in Netlify, decided the casing was wrong, and re-pushed under SlidevCMU. The old repo stayed because deleting a GitHub repo triggers a confirmation dialog that always makes me hesitate. So I have two of them. The disk usage tells the story: cmuSlidev is 3918 KB because I generated a slides-export.pdf and committed it. SlidevCMU is 521 KB because I did not repeat that mistake.
The deck itself is the interesting part. It is a portfolio talk aimed at faculty and internship recruiters at CMU, built on Slidev's seriph theme. Content is the work I did at SAP Labs on the German PII redaction pipeline, the GE Precision Care Challenge submission (SentinelPII, DeBERTa plus QLoRA plus focal loss), and a handful of infrastructure claims. The slide with the numbers reads:
- 94 percent PII recall on German
- 0.82 MRR at 10 for the retrieval layer
- Under 100 ms latency overhead
- 91 percent F1 on multilingual PII for SentinelPII
These are the same numbers I would put on a resume. Slidev's advantage over a static PDF is that the deck is a Vue app, so I can embed a live component next to the numbers. That is what I built the repo for. In practice I never opened it again after the initial push. The talk was hypothetical. I was preparing a pitch for a room I had not entered.
I keep the repos private because private-to-public friction is lower than public-to-embarrassed. If I need this deck in August, it will need a rewrite anyway. Slides written four months before a talk are ghosts of what you thought you would want to say.
The research repo
AIE-ECE-Research is the one I actually spent time in. Created 2026-05-11, last push 2026-06-01, roughly 30 MB on disk. The primary language is listed as JavaScript because there is an index.html plus styles.css plus app.js that renders the data as a TheOrg-style org chart. But the real substance is 14 JSON files.
The files:
msaie_2024_placements.jsonmsaie_2025_placements.jsonmsaie_2026_MASTER.jsonmsaie_2026_profiles.json,_deep.json,_final.jsonclass_2026_complete.jsoncross_year_analysis.jsoncurrent_batch_2027.jsonfeeder_schools_full.jsonlinkedin_data.jsonstudent_career_profiles.jsoninvite_queue.json,invite_remaining.json(the outreach list I never actually sent from)
Plus a resumes/ folder with 29 PDFs from public GitHub Pages of previous students, plus commencement-2024.pdf and four ECE diploma PDFs spanning 2023 to 2026 for the diploma photos.
The scan scripts are scan_resumes.py and scan_resumes_aggressive.py. They are not deployed anywhere. They just read the PDFs and pull out education, roles, dates.
Why does this exist? Because if you are joining a graduate program in three months, the highest signal information you can find is what the people two years ahead of you did. The program brochure tells you what CMU wants you to know. The JSON tells you where people actually went. Amazon SDE, three of them. Uber, two. AWS, two. Google, TikTok, and then the tail: startups I had not heard of, PhD placements at CMU LTI, one confirmed founder (Xue Wang at a company called Xynect AI), one confirmed "still seeking" who was still active on GitHub in May 2026.
The commit history is honest about the messiness. One commit reads "Final research pass: 20/30 Class 2025 outcomes confirmed (67%)". Ten students, one third of the class, I could not confirm. "Kaiwen Lan: confirmed zero web footprint (truly unfindable)" was a real commit message. Absence is a finding.
I used the browser harness against LinkedIn for the disambiguation work. The kai-q LinkedIn slug turned out to be Kai Qiu, not Yichen Lu, which I only figured out because two of the AMD Research intern profiles had different school histories and the AMD one matched Kai's GitHub. This is the kind of work that is either research or stalking depending on how charitable you want to be about the framing.
The feeder_schools_full.json is the file I go back to most. It is a distribution of undergrad institutions for the incoming CMU MS AIE class. It is not for admissions strategy, since I am already admitted. It is for the social geography. If seven people in your class came from IIT Bombay and you did not, that is worth knowing in the first week.
What the reading list looks like from here
The program's suggested pre-reading includes the ORCA scheduling paper (Yu et al., 2022) and the vLLM paper (Kwon et al., 2023). The tone assumes you already know what continuous batching is and what a KV cache eviction policy solves. If you show up in August without those two papers metabolized, you are behind on day one.
So the reading is happening, but it happens next to the strange preparatory work of the two Slidev decks and the AIE-ECE-Research JSON. Which of these will matter in September?
The reading will matter. The slides will not, since I will rewrite them from scratch once I know which faculty are actually accessible. The research repo is the interesting bet. If it turns out that having a mental map of where people from my program end up is useful for career decisions in year one, the 30 MB was worth it. If it does not, I have a JavaScript prototype and a directory of PDFs that quietly ages.
Pre-semester work is preparation without feedback. You cannot verify the slides against faculty reactions, or the reading against a problem set, or the research against a job you have. You build the artifact you can build and let the semester grade it.