Case Study
Course-Tuned AI for Calculus and Physics at AUS
How the American University of Sharjah deployed ibl.ai agents in gateway STEM courses — grounded in faculty-approved sources, able to compute and plot, and measured against published accuracy targets before any rollout.
In math-heavy courses a confident wrong answer is worse than no answer. AUS’s agents for Calculus I and Physics 101 were built to compute and visualize, not just converse — with inline citations back to faculty-approved sources so students can check the work.
Gateway STEM
Calculus I and Physics 101
≥95%
Weekly graphing-accuracy target
≥4.2 / 5
Monthly explanation-quality score
≥70%
Adoption target among enrolled students
Prove it in the hardest courses first
AUS chose two gateway STEM courses — Calculus I (Math 103) and Physics 101 — deliberately. The work was designed to validate instructional impact, technical fit, and day-to-day faculty and student workflows in the courses where accuracy matters most.
Course-specific agents
Each agent carries AUS-specific prompts, tone, and guardrails, grounded in faculty-approved texts and OER, returning inline citations to those sources so learning stays transparent and verifiable.
Cost under the university’s control
Per-student usage controls let AUS manage consumption directly, and they can be adjusted at any point by the university rather than renegotiated with a vendor.
The university’s keys, with a fallback ready
The setup is model-agnostic and runs on AUS-provided API keys by default, with a pre-selected secondary LLM standing by if service quality fluctuates. Switching does not require changing the agents — the model is configuration, not architecture.
To be useful in math, it has to actually compute
The agents run a secure code-execution environment, which is what separates a genuinely useful STEM assistant from a fluent one. Text prediction alone produces plausible-sounding but wrong mathematics.
Plot and render
Precise graphs of equations and vector fields, produced as images students can reference later.
Check work numerically
Verify limits and derivatives, evaluate integrals, and test boundary conditions rather than asserting a result.
Sanity-check symbolic steps
Sample values to spot algebraic slips and compare equivalent forms, catching errors a student would otherwise inherit.
The effect is fewer confidently wrong answers and clearer visual feedback — especially valuable in early calculus and mechanics, where a small algebraic slip propagates through everything after it.
Targets published up front, not after the fact
AUS and ibl.ai committed to concrete numbers and a tight feedback loop — the part most AI programs skip, and the reason most of them cannot say whether they worked.
| Measure | Target | How it’s checked |
|---|---|---|
| Graphing accuracy | ≥95% pass | Weekly 25-item checklist; critical issues resolved within five business days |
| Explanation quality | avg ≥4.2 / 5 | Monthly sampling on correctness, clarity, and alignment to sources |
| Student perception | ≥80% helpful | Students rating the agent “helpful” or “very helpful” |
| Adoption | ≥70% of enrolled students | Unique users, sessions, and messages per session |
Quality is monitored continuously through monthly response audits, targeted spot-checks on graphing, and in-product flagging — faculty, and optionally students, can flag any response. Issues are triaged, fixed with precise prompt or dataset changes, and tracked in a shared log.
Inside the LMS students already use
Access arrives through secure links or LTI, with the technical groundwork — HTTPS, CSP allow-listing, standard LTI claims for user, role, and course — handled up front so access is smooth across browsers and sections.
Faculty enablement
Hands-on working sessions per course for setup, testing, and deployment, plus asynchronous support throughout the term for prompt tuning and dataset adjustments, and an end-of-term debrief on usage, satisfaction, and accuracy.
Run this in your hardest courses
Course-tuned agents grounded in your own materials, computing rather than guessing, on your API keys — with targets agreed before anyone calls it a success.