Case study / Applied AI, own brand
AI a memory-care team uses every day
Planning a month of resident activities that balance seven wellness dimensions takes a coordinator hours and is easy to get wrong. Bloomday does it in minutes, reliably enough to be trusted daily.
In daily production use- Constrained tool-use
- Server-side validation
- Privacy by design
The engagement
Challenge
Trustworthy AI output, for a non-technical daily user
A coordinator plans a month of activities across seven wellness dimensions, slow work, easy to get wrong. The hard part is not generating a schedule; it is making the AI output trustworthy every single time for someone who is not going to debug a bad response.
Action
Fence the generation for reliability
A planner whose AI is constrained by construction: schema-locked tool-use with enum-restricted locations and wellness dimensions so the model cannot invent a value, truncation detection that returns an actionable error instead of garbage, server-side validation and repair of anything invalid, and per-week coverage reporting across the seven dimensions.
Result
Live on the floor, private by design
In daily use on the floor of a real memory-care unit. Privacy is an invariant: the system stores anonymous resident codes, never names or PII, so there is nothing sensitive to leak.
How it is built
Constrain, validate, repair, report
Every generated value is checked and repaired server-side, so a bad model response cannot reach the user.
- 01Enum-locked output. Locations and wellness dimensions are constrained enums in the tool schema, so the model cannot invent one.
- 02Truncation handling. A max-tokens stop is detected and returned as an actionable error, never as half a schedule.
- 03Server-side repair. Invalid or invented values are validated and corrected before anything reaches the coordinator.
- 04Coverage reporting. Per-week coverage across the seven wellness dimensions is computed and surfaced as gaps to fill.
- 05Privacy invariant. Residents are anonymous codes; no names or PII exist in the system.
What it proves
The hard part of shipping AI to a non-technical daily user is not the model. It is making the output trustworthy every single time.
See the rest of the work →Built for a specific memory-care facility. No resident data is shown; the facility is referenced only in the abstract.