Hire on evidence,
not keyword scores.
Refraction is an applicant tracker for small teams hiring interns and juniors. For every requirement you write, it quotes the resume lines that prove it. When a resume leaves something out, it asks the candidate instead of guessing. No score, no ranking, no automatic rejection.
Trained a neural network to predict sleep stages from wearable sensor data, reaching 84% accuracy
Cleaned and joined 2 years of study data in Python and SQL for the lab's weekly analysis
Presented our findings as a poster at the Denman Undergraduate Research Forum
Built a Tableau dashboard tracking weekly sales and inventory, used by the store managers every Monday
Wrote SQL queries to find slow-moving products, which cut food waste by 12%
“Entry level” often isn't.
We read every posting in a public Kaggle dataset of LinkedIn jobs from spring 2024 and counted how many ask for years of experience. Requirements like these screen out the people a junior role is for. Refraction flags them before a job goes live and bends the posting back toward what the work actually needs.
of entry-level postings ask for a number of years of experience.
entry-level postings asks for three years or more.
of internship postings mention a GPA, a cutoff that screens on grades instead of work.
Years of experience asked for in entry-level postings
Share of 36,708 entry-level postings, by the smallest number of years the description asks for.
Postings that ask for years of experience, by level
Share of postings at each LinkedIn experience level whose description asks for a number of years.
Show the numbers as a table
| Years asked (entry level) | Postings | Share |
|---|
| Level | Postings | Ask for years |
|---|
Source: LinkedIn Job Postings (2023–2024) by Arsh Koneru, on Kaggle, licensed CC BY-SA 4.0. 123,849 postings, listed between March 24 and April 20, 2024. Experience level is LinkedIn's own label. A posting counts as asking for experience when its description names a number of years of experience (for example “2+ years of experience” or “1–3 years of related experience”); we checked a random sample by hand. LinkedIn's own analysis of entry-level postings found a similar share, 35%.
From posting to a decision, in five steps.
Each panel shows the actual product, filled in with the same sample candidate.
Plain sentences, not keyword lists. Refraction shows how it reads each one and flags wording that screens out good people.
Post it anywhere. Candidates apply without making an account.
Data science intern
Every requirement gets a verdict and the exact resume lines behind it.
EvidenceHas built a dashboard or data visualization
Tableau is a kind of data visualization, used in a job, where they “built”.
EvidenceExperience with Python and SQL
UnclearHas trained a machine learning model with PyTorch or TensorFlow
Shows deep learning, but doesn't say whether it involved PyTorch or TensorFlow. Not counted either way until the candidate answers.
Vague resume? Refraction asks the candidate, and holds rejection until they answer.
Drag candidates through stages. Every move emails them, including a clear no.
Applied 2
Reviewing 1
Interview 1
Rejected 1
Cards stay in the order people applied. Blind review hides names until the Interview stage.
There's no AI model reading resumes.
Refraction's engine is rules plus a map of how 230 skills relate: PyTorch is a deep-learning framework, Tableau is a kind of data visualization, a poster is research output. It reads the way a careful reviewer does, in about 10 ms a resume, and it can always show its work.
| Keyword match score | AI model screening | Refraction | |
|---|---|---|---|
| Shows why | A number | A summary, which can quote lines that aren't there | The exact lines, highlighted in the resume |
| Same resume twice | Same score | Not guaranteed | Same answer |
| Vague resume | Scores it low | Fills in the gap | Asks the candidate |
| Rejects automatically | Often, below a cutoff | Often, below a cutoff | Never |
| Hidden white text | Counts it | Can follow it | Flags it for a person |
A person makes every decision.
No score, no ranking
Verdicts are Evidence, Partial, Unclear, or Not shown, for each requirement. Cards stay in the order people applied, with blind review if you want it.
Every applicant hears back
No account, an email at every stage including a clear no, and a private page to check their status and delete their data.
A record of who did what
Stage changes, exports, sign-ins, and deletions go into an audit log your team can read. Candidate data is never used to train anything.
Hire your next intern with us.
We're opening Refraction to a handful of teams with a real role to fill. We set it up with you, and you tell us what's missing.
- CostFree for the pilot semester.
- SetupWe help you post your first role and check the requirements together.
- AskA 15-minute call each week, and permission to share what we learn.
- Looking forStartups, research labs, and student organizations in the US hiring interns or juniors.
Questions
What if a resume describes a skill without naming it?
Refraction recognizes related evidence, like a neural network for “machine learning”, and counts it. When the resume doesn't settle a specific requirement, the verdict is Unclear and the candidate gets one targeted question instead of a guess.
What do candidates need?
A name, an email, and a resume as a PDF or Word file. No account. They get a private link to check their status, answer questions, and delete their data.
When can we use it?
Refraction works today and is opening to pilot teams first. Request a spot with a role you're hiring for, and we'll get in touch to set it up.
What will it cost?
Free during the pilot. After that we expect research labs and student organizations to stay free, with a simple monthly plan for startups. We'll set pricing with pilot teams, not before.