CASE NO. 07 · CHILDCARE / VERTICAL SAAS · MBARC VENTURE · IN DEVELOPMENT
Moving kids up a classroom without breaking a friendship or the budget.
ChildcarePlus is an operations platform for childcare centers, rebuilt from a working proof of concept. The part that matters most is how children move up a classroom. They move in small friend-groups, rooms stay near the occupancy they need, and a teacher or a parent signs off before anything happens. An MBARC venture, still in active development.
Built so far
The shape of it
Four services on one operations spine, with a privacy gate standing in front of the model.
The spreadsheet problem
Centers run on tools that keep enrollment, classrooms and transitions in separate spreadsheets. So the decision that matters most gets made by hand, under time pressure, in a document that knows nothing about the room next door. Which children move up, and when.
Why it's hard to get right
Get it wrong and you split up a friend-group, or leave one room half-empty while another sits over capacity. The same call is a developmental question and a social one, and it moves money. That makes it worth optimizing. It also makes it the last decision anyone wants a black box making for them.
The constraint
The model never sees a child's name.
- Names replaced with Child A, Child B before any prompt is built
- Birth dates reduced to age in months
- Medical fields stripped entirely
- A reversible mapping restores real names locally, after the response
- The model explains decisions. It never makes them
Four layers, one decision.
Scoring
How readiness gets scored
Five weighted factors go into a readiness score: age, developmental milestones, teacher assessment, time in the room, and space in the next one. Hard gates run first and can rule a child out before any scoring happens. Directors tune the weights themselves. A center knows its own children better than our defaults do.
Clustering
Cohorts that keep friends together
Leiden community detection runs over a readiness-similarity graph and groups children into cohorts of two to four. The unit stops being one child. It becomes a small group that moves up together.
Optimization
Choosing between good options
A CP-SAT constraint solver weighs every candidate cohort against classroom capacity and staff-to-child ratios. Then it searches for the set of moves that scores best across revenue impact and transition quality. There are usually many arrangements that fit. The solver is there to find the good one.
Inside the build.
Telling a parent why, without telling the model who.
A director has to be able to explain why a child is moving, in plain language rather than in weights and constraints. That argues for a language model. But nothing about a four-year-old belongs in a prompt. So the identity comes out before the request goes out. Names become Child A and Child B, birth dates become ages in months, medical fields are dropped entirely, and a reversible mapping puts the real names back locally once the answer returns. The model writes the explanation. It never makes the decision, and it never receives the data it would need to leak.