Graduate Research Intern — Summarization Faithfulness & Learning Effectiveness

Host: 9keys Group Inc. (WizzStudy) Academic supervisor: TBD (Canadian university faculty in CS, Cognitive Science, or Education Technology) Funding mechanism: Mitacs Accelerate (4-month unit, renewable) Compensation: $15,000 per 4-month term (Mitacs minimum) + co-supervision Location: Hybrid — in-person 1 day/week at the Ontario Tech campus, remote otherwise Start window: Rolling — fall, winter, or summer term

What you'll work on

We're answering one question with a paper trail:

Does our generated study material actually help students remember more in 30 days — and is it traceable to the source material — compared to existing AI study tools?

Concretely, the intern owns one or more of these workstreams:

1. Faithfulness benchmark. Grow the hand-graded corpus from 16 to 200 documents across 8 subject areas. Each entry needs: source text, gold-standard summary, gold flashcards, gold quiz answers, and span-level citation annotations. The existing JSON schema in eval/corpus/ is the format.
2. Eval harness extension. The nightly run in eval/harness.ts currently scores ROUGE-L, BERTScore, and a sampled judge-LLM. Extend it with: claim-level entailment scoring, contradiction detection, and citation-precision/recall. Each addition should be reproducible from a single pnpm eval command.
3. Retention causal analysis. We log FSRS reviews per generated card with a prompt_version tag (see lib/ai/experiments.ts). The intern designs and runs the analysis: do students using variant X show higher 30-day recall than variant Y, controlling for difficulty and exposure?
4. Write-up. Co-author the paper draft in docs/papers/ toward submission to L@S, LAK, or an NLP venue (TACL, EMNLP Findings).

You will not be asked to build product features. That's our job. Your time is for science.

Required background

• Currently enrolled in a Canadian master's or PhD program (Mitacs eligibility).
• Coursework or research experience in one of: NLP, IR, learning sciences, educational data mining, HCI.
• Comfortable with Python or TypeScript (we use TS for the harness; analysis is fine in either).
• Reads papers without flinching. Writes prose that survives review.

Nice-to-have

• Prior work with LLM evaluation (RAGAs, TruLens, judge-LLM frameworks).
• Familiarity with FSRS, SM-2, or cognitive models of memory.
• Experience preparing a Research Ethics Board (REB) submission.
• Exposure to causal inference (propensity scoring, A/B test attribution).

Time commitment

Mitacs Accelerate is a 4-month full-time research unit (~525 hours). We expect:

• 40% measurement work — corpus, harness, analysis.
• 30% writing — drafting, revising, lit review.
• 20% reading group — weekly with the academic supervisor and WizzStudy team.
• 10% slack — for the ideas that don't fit a checklist.

We do not expect a 9–5. We expect the work to ship.

What you get

• Listed first author on at least one paper draft.
• Public attribution on the live eval leaderboard.
• Open-access dataset under your name on Zenodo or HuggingFace.
• Reference letter that says exactly what you did.
• A defensible research artifact for your thesis chapter.

How to apply

Email research@wizzstudy.com with:

1. CV (1 page is fine).
2. One paragraph: which workstream do you want, and why.
3. A link to one piece of writing or code you're proud of.

We reply within 7 days. If your application is shortlisted, we'll do a 45-minute conversation — half on your background, half on a paper we'll share in advance.

We read every application. There is no application portal. There are no tracking pixels in our reply.