I’m Kyle. I design how people and AI work together, and where AI doesn’t belong.

The work below: AI agents that draft courses while experts decide what ships, a guide to when AI fits a task, a planning tool for teachers, and five small agents you can try yourself.

See the work Try the agents Experience Contact

AI course production

United Way Worldwide

Expert review was the slowest part of building a course. When I brought AI agents into course production, I wrote the rules they work under: what they draft from, who reviews their work and who can reject a draft. Every change an expert made went into the agent’s instructions and examples, so drafts came to sound like the expert. The experts were skeptical at first, and over time they made fewer edits to each first draft.

78%: less subject-expert review time per course

Averaged across 11 courses built in 2026.

An AI agent that keeps what the expert changes

Every review goes into the agent’s instructions and examples, so each draft starts closer to what the expert wants. The expert keeps the final say. Used on 11 courses in 2026.

A feedback loop. Frameworks and a procedure Kyle set feed an agent that updates the draft from the expert’s meeting notes. AI checks review it first, the way that expert would. The subject expert reviews every draft and can reject it. Each review goes into the agent’s instructions and examples, not the AI itself, so each round starts closer to the expert and the loop gets smaller. Once the expert approves, the learning team checks quality before the course is published. Expert review time per course fell 78%.

1Frameworks and procedureI set the sources, limits and reviewers
2Agent updates the draftFrom the expert’s meeting notes
3AI checks it firstThe way that expert would
4Subject expert reviewsCan reject any draft
5The agent’s instructions and examples updatedFrom every review. Only the instructions change, not the AI.
6Learning team checks qualityNothing ships without it
7Course published78% less expert review time

One agent per program, including the AI enablement courses.

View the course-production process

Network learning platform

United Way Worldwide

I led the launch of a learning platform and wrote the design standards for its course catalog, in English and Spanish. I set the direction for three contract learning experts and wrote the procedure their work followed. More than 800 of the network’s 1,100+ organizations chose to use it.

How its courses were built

AI curriculum and use guide

United Way Worldwide

I designed a learning path that runs from AI basics to how AI agents work. Learners practice on tasks from their own jobs, using a Green, Yellow, Red guide that asks two questions of every task: what am I giving the tool, and what happens if the answer is wrong? The guide also treats being too careful as a failure: no more than 10% of safe tasks should be blocked. IT and the Office of General Counsel approved the guide.

View the curriculum and guide See the full enablement plan

Leadership development

United Way Worldwide and NuVuX

I led onboarding for new CEOs of local United Ways, coordinating teams across departments. Of the CEOs who completed it, 95% finished a first-60-day action plan.

At NuVuX, I led the development of an executive program in which leaders from companies including Rolls‑Royce and Balfour Beatty worked through a problem in their own organization and developed a prototype.

View the leadership programs

Alf Studio

Independent product

I designed and built AI planning software for teachers, drawing on my experience running design programs in schools. It runs on the same idea as my course agents: the expert stays in charge. Teachers enter a grade, subject and idea, then review and edit a draft teaching unit, and they decide which suggested changes to keep.

Alf was a finalist in MIT Solve’s Accelerate STEAM Education: The 2025 GM Challenge. In October 2026, I gave a talk on the pattern behind it to DC government leaders in the AI track of GW’s executive leadership program.

Alf StudioUnitsHelpDW

Done · 23 secondsGrade 6 · Science · 3 weeks

MS-PS3-3WHST.6-8.7

By mid-afternoon our blacktop is the hottest place on the block, hotter than the park across the street. Why, and what could we plant or build to cool the schoolyard before next summer?

Science · Grade 6 · 3 weeks · Drafted Oct 21

Cooling Our Schoolyard

How can we redesign our schoolyard to cool down our hottest outdoor spaces before next summer?

Edit unit

Saved. 1 change kept.

Week 1Ready

Map surface temperatures across the schoolyard and track how different materials absorb thermal energy.

Conduct the schoolyard surface temperature walk, log readings on varied surfaces, and build the initial class need-to-know board.
· 1 lesson
Run a controlled bench lab measuring how dark asphalt fragments, light concrete, sand, and shaded soil heat under heat lamps over time.
· 2 lessons
Analyze lab data in teams, graph temperature curves, and draw energy transfer diagrams showing solar radiation and surface conduction.
· 2 lessons

Deliverable

A completed thermal data sheet with a labeled energy transfer diagram of the schoolyard blacktop.

Supports

Provide bilingual graphic organizers for lab data tables; offer sentence frames for thermal energy explanations to students with writing accommodations; challenge fast finishers to calculate rate of temperature rise per minute for each material. Give every pair a sentence-starter card: “The hottest spot was ___ because ___.”

Added a scaffold · just now

Week 2Ready

Research cooling strategies and build a scaled prototype shade or greening design to test heat reduction.

Gather facts from local desert plant guides and shade cloth spec sheets to answer self-generated questions about cooling structures.
· 2 lessons
Sketch scaled design options for a chosen hot zone in the schoolyard using selected shade materials and native trees.
· 1 lesson
Build a tabletop scale model of the cooling design and test its interior temperature under a heat lamp against an unshaded control surface.
· 2 lessons

Deliverable

A scaled prototype model with heat-lamp test data showing thermal energy reduction.

Supports

Use visual vocabulary cards for plant types and shade structures; provide partially pre-scaled grid paper for drafting; invite extending students to test two competing shade materials simultaneously and compare cost estimates.

Week 3Ready

Refine the cooling proposal based on peer critique and present the final plan to school leaders.

Run a gallery walk critique where teams review draft cooling models and leave sticky notes on durability, shade coverage, and cost.
· 1 lesson
Revise the physical model or site layout using peer feedback and write a one-page summary explaining how the design stops heat transfer.
· 2 lessons
Present cooling proposals to the school facilities leader and principal, then write an individual reflection on the design process.
· 2 lessons

Deliverable

The finalized cooling proposal poster, scale model, and written design rationale.

Supports

Offer sentence starters for peer critique; allow oral explanation paired with bulleted visual summaries for students with modified output needs; encourage advanced groups to calculate square footage of total shade created at noon.

Critique and revisionReady

Final productReady

Rubric · 3 rows · 4 levelsReady

MaterialsReady

This is one recorded run: a three-week Grade 6 science unit, drafted in 23 seconds and sped up here. See the real run

Is this task safe to give AI?

Ask two things: what am I giving the tool, and what happens if the answer is wrong? Together, the answers decide whether AI touches the task, and who has to approve it.

The use guide

Will AI replace our experts?

Not if every change an expert makes goes back into the AI’s instructions. The expert stays in charge, and has less to fix each round.

Keep the expert in the loop

How much should AI do on its own?

Pick the level by how often the task comes up and whether the result leaves your team. Then give it more room one permission at a time, once testing shows it’s ready.

Five levels of working with AI

Where in our work should AI go?

Draw how the work really runs, including where it waits. Fix what doesn’t need AI. Then put AI where it fits, and a person anywhere a mistake can’t be undone.

Fix the process before adding AI

Experience

United Way Worldwide

Director, Learning & AI Enablement

A network of 1,100 independent local organizations in 36 countries, each with its own board and its own say in what it joins. From 2024 I ran learning and AI for that network: the first shared learning platform and the courses on it, onboarding for new CEOs, course agents that drafted while experts decided, and a use guide that says plainly when AI belongs in a task and when it doesn’t. The argument I made most often was that being too careful is a risk of its own.

NuVuX

Senior Director, Global Learning Programs

Project-based learning programs for schools and companies, taken from Glasgow into Europe, the Middle East and Australia. The part I think about most is Kelvinside Academy in Glasgow, where a program planned without the teachers nearly failed. Training them to run it themselves, and resetting the plan with the school’s leadership, turned a pilot of fewer than sixty students into one every year group took, and three other schools adopted. In 2023 I built early AI co-design tools for educators, the start of what became Alf.

MIT and SCI-Arc

MS in Design and Technology, MIT, on a full merit scholarship. Bachelor of Architecture, SCI-Arc.

Pick an example, or describe a task your team wants to hand to AI. You’ll see how I’d think it through: what I’d need to know, what could go wrong and where I’d want a person to review it. Leave out names and personal details; this site doesn’t save what you type.

Contact

kyle@kylebranchesi.com