AI Triage & Case Management:

Built Dial, an AI triage and case management system that replaced Quo and now runs Emerge's entire student support operation. It handles large percentages of casework autonomously and proactively surfaces the highest need students for case workers to support manually.

Emerge coaches low-income students through CDL training, from the permit exam to a licensed trucking job. Support used to be reactive: a coach helped whoever wrote in that day, and a student who quietly stalled out went unnoticed. Dial scores every student hourly, opens a case the moment someone stalls, and puts them at the top of a coach's queue with the outreach already drafted.

I designed the triage score that decides who gets a coach's attention, combining how far behind a student is falling against peer velocity, how likely they are to respond, and how much coaching time the intervention is worth. I also built the case lifecycle around it: queues split by training stage, close-and-snooze against real dates like a retest eligibility window, automatic time capture on every resolution, and carrier-level delivery confirmation to fix the message failures we had been living with on Quo.

AI Communication & Self-Improvement:

Built the AI communication system that writes the first draft of every message Emerge sends a student, and the weekly loop that teaches it from the edits coaches make.

Coaches were losing their day to context gathering and retyping the same explanations. Now a draft is waiting in their inbox within five minutes of anything a student does, addressed to the channel that student actually responds on. AI is not trustworthy enough to run communications on its own, so nothing sends without a coach approving it.

I split the drafting across four agents so no single model has to be right about everything at once: one looks up the facts and is not allowed to invent a specific we don't have on file, one decides what to say and on which channel, one rewrites it in a coach's voice, and one re-reads the result against our guardrails and fixes what breaks them. I handled the messy edges too, so three texts in a row become one answer instead of three half-answers, call transcripts refresh a pending draft, and a bare “ok” neither triggers a draft nor throws away a good one.

Keeping an AI accurate is normally an endless manual chore, so I made the system improve itself. Every coach edit is recorded, and once a week the system groups a month of them into proposed changes: a new voice example, a new fact for the knowledge base, or a new hard rule. The team votes, tone changes need a majority, and facts and rules need three approvals with any single veto blocking them. Winners go live the following Monday, and the knowledge base mirrors to a read-only sheet so anyone can audit what the AI is allowed to say.

Financial Aid Infrastructure:

Built the financial aid fulfillment system behind Emerge's government-sponsored skilled trades scholarship, which has funded more than 400 students and paid for 281 CDL permits at $76.59 each.

Emerge pays for DOT physicals, permit and license exam fees, DMV fees, and transport costs so that low-income students can clear any financial barrier in the program. A $30 exam fee is enough to end someone's career change before it starts. The system runs two ways: students who can front the cost pay and get reimbursed, and students who can't get the money upfront.

I built the reimbursement side to clear routine claims on auto-approval, so coaches only ever touch an exception. It has paid out $21,523 across 332 students, and because it only pays after a student has actually sat the exam, it never pays a non-completer. That puts a funded permit at $76.59 against a roughly $77K career outcome. I then productized the upfront grant flow so the students who can't pay first aren't locked out of a program built around reimbursement.

I also ran the analysis on my own work, across 190 aid requests, 119 students, and 30,028 user summaries, and used it to argue for cutting scope. Upfront aid costs more per permit ($197 versus $77) precisely because it reaches people reimbursement structurally cannot, so the two are an efficiency-versus-access tradeoff rather than competitors. The receipt-compliance flow was the part that didn't work, asking students to document money they had already spent, so I recommended scaling it back to where the feature actually delivers: unblocking students on adhoc requests.