Question 1

How do you handle FERPA + COPPA when training models on student data?

Accepted Answer

Three patterns. (1) FERPA — student PII (name, ID, DOB, address, email, biometric, schedule) is stripped or pseudonymised before any model touches it; we operate under the school-official exception (34 CFR §99.31(a)(1)(i)(B)) with the district as data controller, a signed DPA, written direct-control language, and no secondary use. (2) COPPA — for any service touching under-13 learners we route consent through the school under the school-consent exception (FTC guidance), keep verifiable consent records, and never serve behavioural advertising. (3) Training inside the customer tenant or a de-identified enclave is the default; raw PII never leaves the FERPA boundary. California SOPIPA and Student Online Personal Information Protection Acts in NY, CT and CO are layered on where applicable.

Question 2

What does LTI 1.3 / LTI Advantage integration with Canvas, Blackboard and Brightspace actually look like?

Accepted Answer

We ship a tool-provider that registers via LTI 1.3 dynamic registration (or static JSON config for tenants that disable dynamic), authenticates with OIDC + signed JWTs, and uses the LTI Advantage services: Names and Role Provisioning Services (NRPS) to fetch the roster, Assignment and Grade Services (AGS) to write scores and rubric line-items back into the gradebook, and Deep Linking 2.0 to let instructors embed activities. Canvas (Instructure), Blackboard Learn Ultra, Brightspace (D2L) and Moodle 4.x are all certified targets. Where LTI 1.1 / Basic LTI is still in play (legacy installs), we ship a 1.1-to-1.3 shim and migrate during rollout.

Question 3

Can you automate IPEDS reporting against our SIS?

Accepted Answer

Yes. We map Banner, Workday Student, PeopleSoft Campus Solutions or Anthology into the 12 NCES IPEDS survey components — Institutional Characteristics, Completions, 12-month Enrollment, Student Financial Aid, Graduation Rates (GR and GR200), Outcome Measures, Admissions, Fall Enrollment, Finance, Human Resources, Academic Libraries — with the CIP 2020 taxonomy applied and edit checks pre-flighted against the NCES rules. Output formats: keypath upload, web XML and the IPEDS Data Center key-holder upload. Title IV cohort default rate, Common Data Set, and state longitudinal data system feeds run off the same warehouse layer.

Question 4

How do you handle AI-text detection now that 2026 student writing is heavily LLM-influenced?

Accepted Answer

We treat binary "AI vs human" classifiers as low-confidence — published false-positive rates against ESL writers and neurodivergent students make them legally fragile and pedagogically harmful. Our default stack is process-evidence plus stylometric drift: keystroke and edit-history signals captured in the LMS, version-history diffs, comparison against the student's prior submissions, and an AI-influence likelihood band rather than a binary verdict. Rubric-aligned feedback runs separately and is the primary product. For institutions that still want classifier output (e.g. Turnitin, GPTZero) we surface it as one signal among several with confidence intervals and never as a stand-alone academic-integrity finding.

Question 5

Are deliverables Section 508 and WCAG 2.1 AA conformant out of the box?

Accepted Answer

Yes. Every UI surface is built against WCAG 2.1 AA and the Revised Section 508 standard (36 CFR Part 1194), with a VPAT 2.5 / Accessibility Conformance Report shipped to the customer. Automated checks (axe-core, Lighthouse) run in CI; manual checks (NVDA, JAWS, VoiceOver, keyboard-only navigation, 200% zoom, prefers-reduced-motion) are part of QA. For content surfaces we ship alt-text, captions (WebVTT), transcripts and adjustable reading-level. State higher-ed systems with stricter accessibility mandates (CSU, SUNY, UC) have their additional checklists merged into the test plan.

Question 6

What is "methodology-faithful RAG" and how is it different from a normal RAG bot for teachers?

Accepted Answer

A generic teacher copilot pulls from open web sources and produces lesson plans that may contradict the publisher's pedagogy. Methodology-faithful RAG retrieves only from a curated, version-controlled knowledge base of the publisher's own strategies, anchor texts, scope-and-sequence, and unit frameworks — with strategy-name attribution on every generation, refusal when retrieval confidence is below threshold, and an editor in the loop for publisher review. In the K-12 writing copilot we shipped, teachers using the system saved approximately 22 hours per week on lesson preparation while producing plans that stayed inside the publisher's methodology. Details and architecture diagram: /case-studies/k12-writing-copilot-22-hours-saved.

Question 7

How do you build a Title IV / financial-aid agent without misadvising students?

Accepted Answer

Three guard-rails. (1) Knowledge grounding: the agent retrieves only from the institution's catalog, Federal Student Aid Handbook (current award year), Title IV regulations (34 CFR Parts 600, 668, 673, 674, 675, 682, 685, 690), and the institution's SAP policy — never from open web. (2) Boundary detection: any question touching dependency status overrides, professional-judgment appeals, eligibility recalculation, or aid-package changes is escalated to a human aid officer; the agent confirms intake but does not advise. (3) Auditability: every conversation is logged with retrieval citations, the rubric used for the boundary decision, and a human-review queue for anything the agent declined. NACUBO and NASFAA guidance on automated communications is followed.

Question 8

How do you prevent bias in early-warning at-risk-student models?

Accepted Answer

Early-warning models trained naively on LMS plus SIS data routinely encode socioeconomic, racial and disability proxies — ZIP code, free-and-reduced-lunch status, IEP flags, English-learner status, attendance patterns. Our protocol: (1) Protected attributes are excluded from features but audited as outcome variables. Suspected proxies (ZIP, schedule density, prior-school flag) are stress-tested with mutual information against the protected set. (2) Training applies adversarial debiasing or constrained optimisation depending on the institutional tolerance. (3) Validation reports subgroup recall, calibration parity, and equal-opportunity difference; a named-disparities document is signed by the institutional research office before deployment. (4) The model output is a flag-for-review, not an automated intervention — the counsellor decides. CRDC and the Department of Education Office for Civil Rights guidance frame the fairness criteria.

AI for K-12, Higher Ed & EdTech

Inference Systems for Education & EdTech

Methodology-faithful RAG for Teachers

Early-warning At-risk Models

Rubric-aligned Writing Feedback

Student-support & Enrolment Agent

Course-pathway & Degree-completion

IPEDS + State Reporting Automation

Learning Data Infrastructure

Learning Data Warehouse

Rostering & Interoperability Layer

Reporting & Institutional Research

FERPA, COPPA, SOPIPA & Section 508 — by design

Frequently Asked Questions

How do you handle FERPA + COPPA when training models on student data?

What does LTI 1.3 / LTI Advantage integration with Canvas, Blackboard and Brightspace look like in practice?

Can you automate IPEDS reporting from our SIS?

How do you handle AI-text detection when 2026 student writing is heavily LLM-influenced?

Are deliverables Section 508 and WCAG 2.1 AA conformant out of the box?

What is methodology-faithful RAG and how is it different from a normal teacher copilot?

How do you build a Title IV / financial-aid agent that does not misadvise students?

How do you prevent bias in early-warning at-risk-student models?

Ship Education AI Without the Compliance Tax