Software engineering services
AI MODERNIZATION
How Incubyte Cuts Campus Hiring Review from Five Weeks to Hours with Claude
When campus hiring brings hundreds of coding submissions, Incubyte uses Claude to score every one against a fixed checklist, so the team spends its interview time on the right candidates.
INDUSTRY
Software engineering services
ON THIS PAGE
The Outcome – Potential Impact
What this could reach once it ships
776
candidate submissions reviewed across five campuses, on its own, in a few hours
194 hours
of manual first-pass review replaced — close to five weeks of work for one person
253
candidates moved forward to interviews — the team spent its time on this shortlist, not reading every repository from scratch
From the client
"Claude gave us a scored first pass in hours, and because it backs every score with evidence, we could trust the shortlist enough to spend our interview time on the right people."
- THE CHALLENGE
When Incubyte hires from college campuses, every candidate submits a coding assignment as a code repository.
Someone then has to open each one, read the code, and decide who is strong enough to interview. But this process is difficult to scale. One campus alone had 198 candidates, and across five campuses there were 776 submissions. Even at 15 minutes each, reading all of them by hand is about 194 hours of work, close to five weeks for one person.
Speed is not the only problem. When rushing through hundreds of repositories, it is easy to miss what matters. A submission can look polished on the surface and still miss half of what the assignment asked for. A candidate can hand in AI-generated code as if they wrote it. A candidate can even fake their commit history to make a last-minute dump look like steady work.
- THE APPROACH
Incubyte built a tool that uses Claude to review each submission.
Before Claude looks at the code, the tool checks the facts it can measure on its own: when the commits happened, how many tests there are, whether the README is complete, and whether any secrets were left in the code.
It then gives Claude those facts along with the code and the assignment checklist, and Claude has to back up every score with that evidence. If Claude rates a submission highly but the evidence shows half the required work is missing, the score is pulled down automatically instead of trusted. Mistakes that were planted on purpose, like a signup that lets anyone make themselves an admin, pull the score down hard. Anything sitting close to a pass-or-fail line goes to a person to review rather than letting the tool guess.
- THE OUTCOME
Across five campuses, the tool reviewed 776 submissions on its own in a few hours.
The same first pass by hand would have taken about 194 hours, close to five weeks for one person. Of the 776, 253 moved forward to an interview, so the team spent its time on that shortlist instead of reading every repository from scratch.
The bigger lesson was that Claude did the heavy lifting while a careful person checked its work, and that pairing caught problems either one alone would have missed.
- STACK
Tech
Claude
Team
Akshay Vadher
Software Craftsperson – Tech Advisor
Sanket Dhokte
Technical Product Craftsperson
Yogendra Jaiswal
Software Craftsperson
Weighing a rewrite?
There’s a third path. Most of our work begins with a thirty-minute conversation.