Tandem

Teachers observe carefully and parents care deeply; the information exists. What's missing is a shared, low-stakes way to reconcile two views of the same child before a concern becomes a crisis.

The same child, two contexts, no shared view. The teacher and the parent each hold half the picture, with no low-stakes way to reconcile them.

The same child ends up living in the gap between two incomplete pictures.

We went wide before proposing anything, pairing AI-assisted synthesis with human judgment: AI summarized and clustered the sources, and I made the interpretation calls by hand.

Each method answered its own question. Click any one to see what we did and what we found.

Across the literature and both interviews, three insights defined what the design had to do:

These three findings pointed straight at a communication tool: one that translates, paces, and aligns.

Research pointed to three barriers: teacher language triggers anxiety, feedback is saved up until it lands as a verdict, and teachers have minutes, not hours. Each became a decision below.

Decision · 01

A guided observe → communicate → act journey

Insight

Communication happens all at once, late, at conferences, so teachers are time-poor and parents feel blindsided.

Decision

Break the exchange into a calm, multi-step journey that mirrors how a concern actually unfolds.

Solution

Nine screens grouped into three acts, each doing a single, low-stakes job:

1. Observe: the teacher logs the concern, frequency, examples, the child's strengths, and what they've tried.
2. Communicate: the AI drafts the message; the teacher reviews and approves, then the parent receives it in plain language, with developmental context and a few questions to answer.
3. Act: the parent adds their home perspective, and both sides build an editable plan together.

Outcome

The journey mirrors how a concern surfaces (notice, talk, act), so neither side carries the whole conversation at once. Later evaluation bore this out: both the real-teacher walkthrough and the synthetic personas gravitated to the structure.

3 acts

Observe Communicate Act

Observe: the teacher logs one structured observation.

Decision · 02

An AI copilot that translates, never diagnoses

Insight

The core problem is translation. Professional language triggers parent anxiety. And in a child-behavior context, an AI that sounds clinical is dangerous.

Decision

Make the AI a copilot that turns structured teacher notes into warm, plain-language messages, with the teacher always in the loop and hard limits on diagnostic language.

Solution

The copilot is the primary pattern; two supporting patterns make it trustworthy:

1. Explainability: a "How this message was created" panel surfaces the observations and patterns behind the draft, so the reasoning is visible, not hidden.
2. Controllability: tone control and regenerate; the model proposes, the teacher decides.
3. Human-in-the-loop: nothing sends unreviewed; every draft is edited and approved first.

The hard rule: Tandem never diagnoses, labels, or names conditions. It translates into strength-based, developmentally grounded language and points to professionals when a concern is serious.

Outcome

This is where the biggest risk lived: an AI that sounds clinical. So the copilot leads with the transparency panel and a bias-aware language safeguard, and the fix was never a smarter model, it was keeping the human in front of it. Our later evaluation put exactly this to the test.

The copilot

The copilot reframes the note into developmental context, citing its source and confidence, and never diagnosing.

Detail · where the safeguards live

The responsible-AI story isn't in the copy: it's in two screens.

Translate, don't dictate. The AI rewrites the teacher's raw note into neutral language, but Accept / Keep Original leaves the final call with the teacher.

Show the work, name the limit. Every AI suggestion cites its source and confidence, and a standing “Not a diagnosis” banner draws the hard line.

Decision · 03

Turn one-way feedback into a shared plan

Insight

Parents and teachers see the child in different worlds and define involvement differently, yet feedback today flows in only one direction.

Decision

Give parents a structured way to add their home perspective, then co-create an editable action plan split across school and home.

Solution

The final act makes the exchange two-way and ends in something concrete:

1. Two-way: parents add what they see at home through focused prompts, enough to balance the picture without becoming a form.
2. Aligned action: an editable plan split into school-side and home-side steps, so responsibility is shared.
3. Follow-through: lightweight reminders and check-ins keep the plan alive past the first conversation.

Outcome

The exchange becomes collaborative instead of one-directional. In our later evaluation, the teacher called the shared action plan the most valuable part: "it focuses on solutions together instead of only pointing out problems."

Two-way

Parents add what they see at home through focused prompts: the first time feedback flows both ways.

Together, the three moves turn a twice-a-year verdict into a continuous, low-stakes loop (observe, translate, align), with the child, not the tool, at the center.

Tandem didn't ship. We evaluated it with three research-grounded synthetic personas (built from our interviews, run as AI-simulated users), an AI-assisted heuristic & risk review, and a think-aloud with one real teacher. These are design-validation signals, not measured outcomes.

If it shipped: the outcomes the design targets

Intended outcomes the design targets, useful for surfacing edge cases, but no substitute for real classrooms.

Our evaluation leaned heavily on AI-simulated users. The synthetic personas surfaced real edge cases, but their reactions came out tidy and a little theatrical; prompt wording alone could swing them from glowing to harsh. I treated their output as hypotheses, not findings, and grounded them against a real teacher walkthrough. The concept itself is unvalidated in a real classroom.

Tandem didn't ship, but it changed how I design with AI: the model earns trust by staying a draft, not an authority.