How an EdTech Platform Designed Its AI to Use 90% Less Input

At this education technology company, notifications ran on a legacy system where every product team sent messages its own way, with no central control or prioritization. Urgent messages and routine nudges reached students with equal weight, and the ones that mattered got lost. Students exploring college and career options saw generic pages with no evidence-backed explanation of why a path fit them. The company wanted to bring AI deeper into that experience, and for young people making real decisions about their futures, doing it responsibly and affordably was a hard bar to clear.

The work spans two builds. The first is a centralized notification platform. It pulls every product team's requests into one pipeline, a prioritization engine scores each message by urgency and user preference, and a scheduler delivers it to mobile or email at the right moment, so the messages that matter surface and the rest stop piling up. The second is an explainable AI personalization layer for college and career pages. For each student it generates a plain-English "Why For Me" narrative and "Fit Signals" cards, with every claim tied to a specific data point from the student's profile, drawn through Model Context Protocol. The defining decision is what the AI isn't allowed to do. Data retrieval, scoring, selection and validation run in conventional software on Azure Functions, and GPT-4 through Azure OpenAI only writes the final student-facing language. The backend pre-scores the relevant options, packs only the student responses tied to each one into a compact evidence packet, and sends a tightly controlled prompt. The model returns structured JSON that is validated against schema and writing rules before it reaches a student. Because the logic lives outside the model, the client can evaluate or swap models without rebuilding the product.

The engagement began as a rebuild of the company's notification infrastructure. Every product team had been sending messages its own way, so the first build pulled those requests into a single pipeline, added a prioritization engine that scores each message by urgency and user preference, and a scheduler that delivers it to mobile or email at the right moment. The second build added an explainable personalization layer to the college and career pages. The team set one rule early: code handles retrieval, scoring, selection, validation and caching on Azure Functions, and GPT-4 through Azure OpenAI only writes the final student-facing copy. The backend pre-scores options, compresses them into evidence packets holding only the relevant student responses, and sends a tightly controlled prompt. The model returns JSON that is validated against schema and writing rules, with retry and repair when live data is messy or incomplete. Before any production code, a batch run wrote 120 real outputs to a spreadsheet for leadership to review. Midway through, a requirement to keep student data inside the client's Azure environment shifted model selection from the cheapest external provider to the safest compliant option. First meaningful results arrived in four to six months.

Edtech and learning platforms, companies under strict data-residency commitments, and multi-product SaaS organizations.