Research internships, fellowships, student researcher roles, AI residencies and predoctoral positions at AI labs, big tech and startups — collected every day, with only the new ones flagged.

→ dion-jy.github.io/phd-intern-board

Built because the existing aggregators are dominated by software engineering internships, and the one list that was actually research-focused (frankaging/awesome-ai-research-intern-list) has been dead since 2024. Research postings also tend to close within days of appearing, which makes checking 70+ career pages by hand unsustainable.

No HTML scraping. Every source is a public JSON endpoint that needs no authentication, which means there is no layout to break and no robots.txt to worry about.

Listings are categorised on the site by kind of organisation — frontier lab, big tech, startup, infra & chips, research org — because that is what a reader actually filters on.

Separately, each listing carries a kind recording how it was collected. That matters for judging what might be missing, not for reading any single listing, so it appears only as a footnote on the card:

1 — Never decide a board exists by counting items. Lever's failure response is {"ok": false, "error": "Document not found"}. That is a dict, so len() returns 2, and a naive check reports two openings for a company that has no board. mistral-ai, together-ai, cohere and adept were all recorded as live that way before it was caught. probe_slugs.py decides on response shape instead: Lever must return a list, Greenhouse a dict with a jobs array, and Ashby anything other than the plain text Not Found.

2 — A slug can respond correctly and belong to somebody else. Caught by verify_identity.py, which checks the owning company name rather than just the payload:

Without this second pass, three companies would have quietly pushed unrelated postings into the feed every morning.

Slugs cannot be guessed. OpenAI is not on Greenhouse but is on Ashby. AI2's slug turned out to be thealleninstitute after seven reasonable guesses — allenai, ai2, alleninstitute and others — all failed. It was found by grepping the careers page HTML, which is the reliable method.

Everything below is a rule that exists because something specific went wrong.

• Word boundaries are mandatory. A substring search for intern also matches Internal Controls, International Tax and Internal Communications.
• Structured signals beat regex. Ashby exposes employmentType: "Intern" and Lever exposes categories.commitment: "Intern". Both are used ahead of any title matching.
• Anthropic never uses the word "intern". Of 518 openings, the research-internship equivalent is the Anthropic Fellows Program — 4 roles, none of which contain intern. The keyword list therefore has to include Fellows, Residency, Scholar and Visiting Researcher.
• AI2 does not use it either. Its pre-PhD position is called Predoctoral Young Investigator, matching neither intern nor fellowship.
• A named programme form is a signal on its own. A fellowship, residency or scholarship is the research-track role, so requiring a separate word like "research" alongside it drops real openings — Scale AI's STEM Fellow and SWE Fellow, Tenstorrent's CPU Verification Fellow. These are kept at medium because nothing in the title confirms the subject area, and the blocklist still removes Scale AI's Finance Fellow and Legal Fellow.
• residency is never matched alone. On its own it hits data residency in infrastructure job descriptions. It always requires a qualifier.
• "PhD appears near research" is far too weak a body-text rule. Etched's chip internships say Bachelor's, Master's, or PhD degree in electrical engineering and carry the boilerplate we do not have boundaries between engineering and research. Both signals are incidental, and nine hardware roles came through. The rule now requires a phrase that names a research position, which cut Etched from 10 matches to 1.
• Discovery rows get filtered twice. Simplify's own AI/ML/Data category still admits Capital Markets Intern - Quantitative Strategies, so discovery titles also go through the ATS filter — 127 down to 92. Without the category gate first, quant trading and defence dominate: of ~400 active PhD listings, ~50 are categorised Quant outright, and Citadel, DRW, Optiver, Tower Research and L3Harris post plenty more under Software and Hardware.
• Blind-spot rows skip that second pass, because NVIDIA files genuine research roles under Hardware and those labs are already trusted.

Small research labs frequently post no internship at all and hire through a standing "general application" or "expression of interest" instead — which makes that channel the only real way in at precisely the labs this board exists to cover. Those carry no internship signal and no date worth trusting, so they are collected as their own kind rather than mixed into the dated postings: no relative date is shown, and they are tagged no deadline.

11 are currently live, including Anthropic's [Expression of Interest] Research Engineer / Scientist, plus standing applications at Goodfire, METR, Epoch AI, Reflection AI, Prime Intellect and Cognition. Channels that cannot lead to a research role are filtered out — Apollo Research's finance associate opening, Scale AI's hackathon interest form and Goodfire's senior-leadership form are all dropped.

Each listing carries a confidence:

• high — the title alone settles it (Research Intern, Anthropic Fellows Program)
• medium — an internship signal plus a research signal, but not conclusive. Kept and tagged rather than dropped, so the site can filter it out.

The one gap automated collection cannot close: labs that announce an internship in an X post, a mailing list or a line on their own site, and never file it anywhere machine-readable. Sakana AI and AMI Labs are the usual examples. Every automated route into X is shut — the v2 API needs a paid tier to search, the syndication endpoint rate-limits immediately, and Nitter is dead — so these are listed directly and trusted without filtering.

Append to manual.yaml. Only company, title and url are required:

openings:
  - company: Sakana AI
    title: Research Intern, 2027
    url: https://x.com/SakanaAILabs/status/1234567890
    via: x                # x | email | referral | lab-page | conference | other
    tier: frontier        # frontier | research-org | bigtech | startup | infra
    posted: 2026-08-20    # when it was announced, not when you added it
    location: Tokyo, Japan
    phd: true
    deadline: 2026-09-30
    note: Apply by DM; mentioned in a thread reply.

A malformed entry is skipped with a message rather than failing the run, so a typo cannot break the daily collection. note and deadline are worth filling in — an X post often carries the only application instructions there are, and the post itself may be deleted later.

There is no automatic expiry: delete an entry once it closes.

Listing schema:

{
  "company": "Waabi", "tier": "startup", "ats": "lever",
  "source": "ats", "kind": "tracked",
  "title": "2026 Intern, PhD Research Scientist",
  "location": "Toronto", "url": "https://jobs.lever.co/waabi/...",
  "posted_at": "2026-08-19T...", "department": "Internships / Co-ops",
  "confidence": "high", "match": ["ats-intern-flag", "title-strong"],
  "phd": true, "job_id": "lever:waabi:62700386-..."
}

• Small research labs post nothing at all. Sakana AI, AMI Labs and similar hire through X, faculty referrals and personal email. They are outside the reach of any automated collector and have to be tracked by following people.
• Big tech is only partially covered. Meta, Apple, Microsoft, Amazon, NVIDIA and ByteDance run their own applicant systems. Google DeepMind does have a Greenhouse board, but it holds ~10 roles while the real Student Researcher postings live on google.com/about/careers. The blindspot rows fill some of this gap through SimplifyJobs — they are not exhaustive.
• Zero results is not the same as "not hiring". Research internships are seasonal; on 2026-08-23, OpenAI had 753 open roles and none of them were internships. That is the reason for a daily diff rather than a one-off scrape.
• Degree requirements, visa sponsorship and start dates are frequently ambiguous in the posting text. Always open the original.

pip install requests pyyaml
python collect.py       # -> api/*.json
python build_site.py    # -> data.js, index.html
python -m http.server   # then open http://localhost:8000

.github/workflows/daily.yml runs both scripts at 01:00 UTC (10:00 KST) and commits only when something actually changed.

This repository sends nothing. A separate agent reads api/new.json on a schedule and composes the briefing. Note the first_run flag: on a cold start every listing counts as new, and a briefing should suppress that case.

index.html mixes hand-written source with regions build_site.py rewrites (the SEO description, the pre-rendered rows, the ItemList, the logo CSS, the feed count). The daily workflow commits it, so rebasing a local change onto a daily commit conflicts on it every time.

Resolve it with --theirs, i.e. keep the local version, then re-run build_site.py to regenerate the generated regions. Taking --ours looks like the safe choice for a generated file and is wrong here: it discards the hand-written half. That has already silently dropped a ordered() helper once and the Open Graph tags once.

for f in $(git diff --name-only --diff-filter=U); do
  if [ "$f" = "index.html" ]; then git checkout --theirs -- "$f"; else git checkout --ours -- "$f"; fi
  git add "$f"
done
git rebase --continue
python collect.py && python build_site.py