AI Coach for EdTech

Context

CodePath.org provides free computer science education to students who'd otherwise pay for it. I led design as a client engagement across 2025 and 2026, through a period when the organization's own product thesis was shifting: CodePath launched a partnership with Anthropic, and today runs a 12-month fully paid program placing early-career AI-native builders inside nonprofits.

Where it started, and where it went

The brief was a gamified student hub — progress, streaks, achievements, the standard motivational scaffolding.

It didn't survive contact with the problem. Gamification answers "how do I keep you here," and the students' actual question was "am I building toward the job I want." Those need different products. Points reward activity; what students needed was guidance on which activity was worth their time, given where they were and where they were trying to get.

So the hub became an AI coach. Recommendations grounded in three inputs — the student's career goals, their performance history, and their personal context — so that guidance was specific to them rather than generic to the curriculum.

The hard problem: a tutor that won't answer

The second system was harder. Alongside the IDE, a chat tutor that helps students during assignments.

Every instinct in AI product design points the wrong way here. A good assistant is fast, complete, and gives you what you asked for. A good tutor is none of those things. A student stuck on a loop can get working code from any model in ten seconds — and learns nothing, then fails the interview six months later. The product's job was to sit in the gap between "stuck" and "told," which is exactly where learning happens and exactly where the model wants to short-circuit.

That made the design work mostly about restraint. The tutor never returns runnable solution code. It opens by asking the student what they think is happening, so the first move belongs to them. Help escalates in steps — reframe the problem, point at the relevant concept, narrow to the line — and each step requires the student to try again before the next one unlocks. A student who asks outright for the answer gets the next rung of the ladder, not the answer.

The escalation ladder was the whole design. It gave us a single place to tune how much help was too much, and it made the tutor's behavior legible to instructors, who needed to trust that the tool wasn't quietly doing the assignment.

The rest of the surface

The Anthropic collaboration drove a redesign of the hub and its sub-pages, and I worked the product side of repositioning CodePath around that partnership. I also designed experimental surfaces: an interview coach, and admin tooling for the schools running the program — three audiences with genuinely different stakes on the same platform.

Outcome

The engagement produced a redesigned student hub, a working coaching system, and a tutor model that gave CodePath a defensible position on AI in education at exactly the moment the sector needed one — a way to put a model in front of students without undermining the reason they're there.

The escalation ladder is the piece I'd carry into another product. It turned an unbounded question — how much should AI help — into a tunable one.

What I took from it

The useful version of AI in a product is rarely the fastest one. Here the entire design problem was deciding what the model shouldn't do, and that constraint came from understanding what the student was for, not from a question about model capability.

It's the same instinct I brought to fraud investigation at Sift, where AI surfaced context and evidence but never made the call. Both times the interesting work was drawing the line, not crossing it.