Neighbor News
Dougherty Valley Junior Built a Public Benchmark to Test How Safe AI Tutors Are for AP Computer Science
Kaushik Atla's open project EduSafeBench scores AI coding assistants on correctness, teaching quality, and safety—using reproducible public
My name is Kaushik Atla, and I’m a junior at Dougherty Valley High School in San Ramon. Like a lot of students, I’ve watched classmates lean on AI coding assistants for AP Computer Science work—especially debugging, explanations, and “is this right?” checks.
That convenience comes with a real problem: it’s hard to know which tools are actually trustworthy for learning, not just “good enough to output code,” but accurate, pedagogically sound, and safe for beginners.
So I built EduSafeBench: an open-source reliability benchmark focused on AP CSA / AP CSP-style scenarios. It evaluates assistants across four dimensions:
Find out what's happening in San Ramonfor free with the latest updates from Patch.
- factual correctness
- pedagogy quality
- hallucination risk
- unsafe guidance risk
Everything is public and reproducible:
- Website: https://kaushikatla-cell.github.io/EduSafeBench/
- GitHub (code + datasets + results): https://github.com/kaushikatla-cell/EduSafeBench
What’s in the project today
Find out what's happening in San Ramonfor free with the latest updates from Patch.
- A 300-item cited benchmark set (AP-aligned prompts with source citations)
- Multi-model leaderboard artifacts you can open directly
- A hard misconception stress-test slice aimed at common beginner traps
- A structured reviewer submission template so feedback is consistent and usable
Why I’m posting this locally I want feedback from people who actually see AP CS classrooms day-to-day: teachers, tutors, club mentors, and strong student mentors. The goal isn’t “hype AI.” The goal is transparent measurement so families and educators can make better decisions as these tools become default study aids.
How you can help (10–20 minutes)
- Skim the website + one report linked from the homepage.
- Use the reviewer template here: docs/reviewer_submission_template.md (in the GitHub repo).
- Tell me what’s missing, what’s misleading, and what would make this genuinely useful in a real classroom.
- Optional but high-value: whether this would influence tool choice or classroom policy for you—and why.
Disclosure I’m the student author/maintainer. Links are public GitHub/GitHub Pages resources with no paywall.
