The platform
The expensive part is already built.
Most of what makes caregiving AI hard has nothing to do with any particular condition. Memory, safety, tone, cost, accessibility, the discipline to say “I don’t know” — those are the same problems whether the person you love has autism or dementia.
The split
A generic harness, and a deliberately thin domain layer.
We built KeyAide so the two would come apart cleanly. Not as a growth strategy — because it is the only honest way to hold the line that domain expertise cannot be faked, while refusing to rebuild a memory pipeline every time.
The harness
Ports as-is
Nothing here is specific to neurodiversity. It is what any caregiving context needs before it can have a single useful conversation.
- Streaming conversation engine
- Multi-model routing across providers and tiers
- Per-conversation cost accounting and budget alerting
- Consent-gated memory: extraction, decomposition, embedding, pattern detection
- Embedding search over a caregiver’s own history
- Guided-conversation engine for structured tool flows
- Care playbook generator with audience-specific output
- Accessibility system, audited against WCAG 2.1 A/AA
- Trauma-informed interaction patterns
- Clinical tone and safety eval harness
The domain layer
Re-authored, every time
This is where the clinical judgment lives, and it does not transfer. It is written fresh for each context, with someone from inside it.
- The system prompt stack — where clinical judgment lives
- Guided-conversation configs for each domain tool
- Playbook content and the audiences it is written for
- Domain-specific tools (an IEP translator becomes a benefits-letter decoder)
- The trusted-source allowlist for that field
- The affirming-stance equivalent for that community
Vertical one
What building KeyAide actually taught us.
Caregivers of neurodivergent kids sit where education law, clinical care, behavioural support and family logistics meet — four systems that don’t talk to each other. We chose it because if the approach works there, it travels. Four things we didn’t know at the start:
The domain layer is smaller than it looks, and harder than it looks
By volume of code it is a fraction of the system. By volume of judgment it is nearly all of it. Every hour saved on infrastructure went back into the prompt stack, the guides, and the question of what not to recommend.
Tone is a testable property, or it drifts
We learned this the expensive way. A voice held together by good intentions degrades every time a prompt is edited. The eval harness exists because the alternative is discovering a regression through a caregiver having a bad day.
The community veto is a feature, not a constraint
Filtering out approaches the autistic community has documented as harmful meant rejecting some material with an evidence base behind it. That choice is the reason the tool is trusted, and it has to be made fresh in every domain.
Accessibility is cheapest at the start
Building to WCAG 2.1 AA from day one cost us very little. Retrofitting it, as the organisations facing that deadline can tell you, costs a great deal. Any vertical we build inherits this for free.
The honest version
A new vertical is not a system prompt away.
The prompt stack is the smallest artifact by keystroke and the largest by judgment. Each domain needs its own clinical validation, its own understanding of what the affected community has flagged as harmful, and its own hard-won trust — which transfers slowly, if at all. Anyone who tells you otherwise is selling a chatbot with a new name on it.
What we can say plainly: we have done this once, for the hardest household we could find, and the infrastructure underneath it did not need to know which condition it was serving. That is the claim. It is not the same as a shipped product in your field, and we will not pretend it is.
How a build actually goes
Four things, in this order.
The sequence matters more than the timeline. Getting it wrong is how well-meaning care technology ends up unused.
Together
Find the domain advisor first
Before anything is written, we need someone with real standing in that community and real clinical or practice judgment. Not a reviewer at the end — a co-author from the start. If we can’t find that person, we say no.
Us
Author the domain layer
The prompt stack, the guides, the tools, the source allowlist. This is the bulk of the work and it is deliberately the part that cannot be shortcut — it is where the judgment concentrates.
Together
Build the eval before the launch
A golden set of scenarios for the new domain — its crisis-adjacent moments, its boundary questions, its known-harmful advice to filter out. Nothing reaches a caregiver until it passes.
You
Reach the people it was built for
The part we are worst at alone. A partner who already holds trust in a community is the difference between a good tool and a used one.
Thinking about a caregiving context we haven’t built for?
We’d rather hear the problem than pitch a solution. Tell us who you serve and what breaks down for them.
Whenever you’re ready.
KeyAide is free. No login wall, no medical claims, no judgment. Just a careful voice in your corner when you need one.
Walking the labyrinth, together