Skip to main content
Gervin Kahunde
Prototype · synthetic data

My Path: AI Student Retention Agent

Online students often withdraw over small, fixable barriers — a hold, a small balance, a missing aid document — that no one spots before the drop date.

  • AI and automation
  • Product Operations
  • Quality assurance
  • Strategy

AI-assisted step Human decision Observed friction

Summary

Challenge
Help success coaches reach students with small, fixable barriers before the drop date, without letting AI decide anything about a student.
Role
I own it end to end: I wrote the PRD, designed the rules and guardrails, and built the API, coach interface, tests, and deployment.
Team
Hack-day concept with my team; prototype built independently for ASU Online success coaches
Duration
2026, in active development
Methods
PRD writing, rules design, responsible-AI guardrails, accessibility testing, CI/CD and SRE practices
Outcome
Live prototype using synthetic data only. Not deployed with real students; no real messages are sent.

This is a working prototype. It uses synthetic data only, sends no real messages, and has not been deployed with real students.

Context and problem

Many withdrawals start with problems that are small and solvable: a registration hold, a modest balance, a failed payment, a missing financial-aid document. Coaches can fix most of them in one conversation — if they know in time. With large caseloads and data spread across systems, they often don't.

My Path gives coaches a short, urgency-sorted list of students with a specific barrier and a clear next step, well before the drop date. The idea began as my pitch at an EdPlus hack day, framed as a synthetic-data MVP with a one-term pilot goal, and was reworked with my hack-day team after feedback.

My role

After the hack day, I took the concept forward on my own. I wrote the product requirements, defined the barrier rules and the AI guardrails, designed the coach workflow, and built and shipped the prototype, including its tests, accessibility checks, and deployment pipeline.

Why AI, and where not

Rules decide
Six deterministic, unit-tested rules decide which students are flagged: small balance, registration hold, failed payment, missing aid document, no recent activity, and not registered for next term. Thresholds live in configuration, so they can be reviewed and changed without code.
AI only explains
The model sees only the flagged fields and returns a plain-language explanation and a draft message. It never decides eligibility, money, or whether a student is contacted.
People approve
Nothing is sent automatically. A coach approves, edits, routes, or dismisses every draft.

Data and privacy boundaries

  • Synthetic data only: the prototype uses a generated set of 200 students and sends no real messages.
  • The AI receives only the fields behind a flag, and every draft must cite only those fields.
  • For a real pilot, drafting would point to an institution-approved AI endpoint that doesn't retain data.

Agent workflow

  1. Step 01

    Upload records

    CSV checked for required columns

  2. Step 02

    Rules flag barriers

    Six tested rules, plus days to drop

  3. Step 03 AI-assisted

    AI explains + drafts

    Guardrails, with template fallback

  4. Step 04

    Coach queue

    Sorted by urgency, filterable

  5. Step 05 Human decision

    Coach decides

    Approve, edit, route, or dismiss

  6. Step 06 Human decision

    Route

    Bursar or financial-aid office when needed

Rules decide, AI explains, the coach acts.

Guardrails and human review

Every AI draft must pass automatic checks before a coach sees it:

  • Under 120 words, at a grade-8 reading level or lower.
  • Greets the student by name and asks for one action.
  • No shaming or risk language.
  • Cites only the fields that caused the flag.

If a draft fails any check — or the model refuses, times out, or errors — My Path falls back to a reviewed template, and the fallback is counted and alerted on. Coaches can dismiss a flag with a reason, undo actions, and redraft in English or Spanish. A full action log records every change, and record locking stops two coaches from overwriting each other.

Quality, accessibility, and operations

Detection
Rule tests check recall on deliberately seeded barriers, which is tested at 100%.
Accessibility
Mobile-first interface built to WCAG 2.2 AA, with automated axe scans on every page and end-to-end tests on phone and desktop.
Delivery
Every change passes linting, strict type checks, unit and end-to-end tests, and security scans before deploying to staging and production.
Operations
Service-level objectives, alerting, a runbook, and an incident-response process are documented alongside the code.

Evaluation criteria for a pilot

Usefulness
Share of flags coaches act on rather than dismiss.
Timeliness
How far before the drop date students are reached.
Fairness
Flag and outcome rates compared across student groups.
Coach effort
Time to review and act on a flag; how often drafts need editing.

Pilot design and risks

A pilot would start with a small group of coaches working alongside current practice, with clear stopping criteria. The main risks are false alarms that waste coach time, missed students that create false confidence, and biased signals. The design answers each one: transparent rules with configurable thresholds, coach dismissal reasons as feedback, and regular fairness reviews.

What I learned

Separating the decision from the explanation made the whole system easier to trust and to test. Because rules decide and the model only drafts, I could guarantee behavior with ordinary tests, and the AI became an improvement that can fail safely rather than a risk.

What's next

  • Usability sessions with success coaches on the review queue and draft editor.
  • A small, approved pilot plan with privacy review and an institution-approved AI endpoint.

Continue with work that uses a similar approach.

In use

Credit Cruncher & TransferMe

Working out which transfer credits will count meant reading transcripts by hand, one course at a time — slow for staff and opaque for students.

My role
Product support: QA, walkthroughs, user stories, success metrics, design sessions, and a prototype
Evidence
Helped take an internal AI-assisted tool through QA and documentation, and helped shape its student-facing successor.
  • AI and automation
  • Internal tools
  • Quality assurance
  • Product Operations