Miami Dade College — STEM PACTS

The problem

STEM PACTS — the Program for Advancing Career & Technical Skills in Miami Dade College's School of Science — supports Title V / Title III HSI grant operations and serves students, tutors, advisors, and faculty across multiple campuses. Its data collection ran on spreadsheets: tutoring sessions, student bookings, event attendance, advising notes, and lab usage lived in separate files maintained by different people.

That made the day-to-day harder and the federal reporting painful. Every cycle for the Annual Performance Report (APR) and external evaluators meant weeks of manual aggregation, with no reliable way to tie an interaction back to a verified student, cohort, and term — and no way to see progress in real time.

The approach

I started where the work actually happens — with the people entering the data — and designed around their day rather than an idealized process. The goal was a single, auditable source of truth that staff would use, where every interaction is logged, time-stamped, and tied to a verified user, cohort, and term, so reporting becomes a by-product instead of a separate project.

What I built

A full-stack platform that consolidates the program's operations, compliance, and reporting:

• Operational tools — tutor session logging, a calendar-first student booking system, event management with check-in and email confirmations, the REEFS lab with laptop/device checkout, advising logs, and the Shark PALS peer-mentoring program.
• Compliance & data integrity — role-based access control, parameterized queries, audit trails, and standardized terms and cohorts so the data holds up to evaluator review.
• Reporting & analytics — APR metrics, cohort and retention analysis, and drill-down dashboards that track each grant objective against target.
• An AI layer — Gemini-powered weekly insights, automated narrative generation for reports, student-risk dashboards, and "STEM PACTS Assist," an internal admin chatbot over the program's own data.

Governed AI access (2026)

In 2026 the program asked a harder question: could authorized staff query the platform from an AI assistant without the assistant becoming a side door around the permissions the platform already enforces? The data is FERPA-protected student information, so the answer had to be provable, not promised. I built a read-only remote Model Context Protocol (MCP) connector around one rule: the assistant inherits the user's permissions, never its own.

• Three scope tiers, per user — aggregates only (for external evaluators), individual records, and ad-hoc query (for administrators). Nobody gets a broader view through the assistant than they have in the platform.
• Every query is logged to a person — not to a shared service account — so an evaluator can see who asked what, and when.
• Leaked credentials are refused — a higher-privilege credential presented in a way that leaks (a URL, a screen share) is rejected rather than honored.
• Read-only by design — the connector can answer; it cannot change, delete, or bulk-export the record.

The same test applies to any AI option an organization is weighing: do the permissions configured in the system of record survive retrieval, indexing, processing, and output? Here they do, and it's verifiable in the logs.

The result

Weeks of manual aggregation collapsed into a live dashboard, and federal reporting is generated from clean, verified data rather than rebuilt by hand each cycle. The AI layer turns that same data into weekly insight and narrative drafts, so staff spend less time compiling and more time acting. Most telling: the team uses it every day across campuses — the mark of software that fits the work instead of fighting it.