A phone call is a hard engineering problem.
Eight kilohertz of bandwidth. A hearing aid whistling. A television on in the background. Someone who takes four seconds to start a sentence and does not want to be interrupted.
Most voice AI is demonstrated on a laptop microphone in a quiet room. None of the people we call are in that room.
One turn of one call
What happens between someone speaking and hearing an answer.
Hearing
Speech recognition tuned for the telephone band rather than a studio microphone, running on GPU so the transcript arrives while the sentence is still warm. Silence and noise get misrecognised as words by every system on the market; ours detects that pattern and refuses to act on it rather than answering a question nobody asked.
Thinking
A language model does the talking, but it does not decide what the turn is about. A priority ladder does — repair first, then an outstanding request, then humour, then following up on something the person just said. The model is handed one job at a time. Models drift when you ask them to hold six goals at once; this is the mitigation.
Speaking
Voice synthesis with one consistent voice, because a companion whose voice changes between calls is not a companion. Generated audio is cleaned before it is sent — synthesis models emit a burst of high-frequency noise at clip onset, which is inaudible on a laptop and unpleasant in a hearing aid.
Repairing
When the system does not understand, it does not say "I didn't catch that" three times. A three-rung ladder makes the question lighter, then easier, then takes the blame itself and changes the subject. Nobody should have to feel tested by a telephone.
The part not allowed to be clever
Escalation is separate from conversation and runs on every turn.
It is not a summary written afterwards and it is not a judgement call made by a language model. It sorts into tiers, and the tier decides who hears about it and how fast.
| Tier | What it is | Where it goes |
|---|---|---|
| Emergency | Chest pain, a fall, a stated intent to harm themselves | On-duty staff, immediately, mid-call |
| Clinical concern | New pain, confusion, a medication problem, not eating | On-duty staff, same shift |
| Safeguarding | Abuse, neglect, or financial exploitation | A designated administrator only — never the floor, never the shared board, never automatically the family view. A person decides what happens next. |
| Wellbeing signal | Low mood, isolation, a slow change over days | The morning board, ranked |
| Request | A leaking tap, a missing hearing aid battery | The right department |
Why safeguarding gets its own route. If a resident says something about a member of staff, and that goes to the shared board that staff can see, you have not built a safeguarding system — you have built a way to get someone hurt. Tier 2 acknowledges and stops. It does not investigate, it never promises secrecy, it writes a verbatim write-once record, and that record is kept in a separate restricted store for five years, because California allows a financial elder abuse claim to be brought within four years of discovery.
And not the family, automatically — which is the part that sounds wrong until you look at who the perpetrator usually is. In the most commonly reported form of elder abuse, financial exploitation, the alleged perpetrator is more often a family member than a stranger. A dashboard that lit up the moment a resident disclosed would, in some share of cases, be telling the accused that they had been named. That is how people get pressured, isolated, or cut off from the person they confided in.
So the software does not make that call. It routes to one named human, and that human decides — including whether to tell the family, which is often the right thing to do. In a licensed California community that person is typically a mandated reporter under Welfare & Institutions Code § 15630, whose duty runs to the authorities the statute names, on the clock the statute sets. Not to us. Not to a screen.
The resident also gets a say. A capacitated adult who discloses something has the right to decide who else knows, and an automatic notification takes that from them at the worst possible moment.
Everything else stays a signal rather than a transcript. Staff get who to see first, not a pile of reading.
What is running, and what is not
Operators and grant reviewers both deserve to know which half is working software.
| Component | State | Detail |
|---|---|---|
| Voice stack, end to end | running | Recognition, conversation engine and synthesis run as a live pipeline on development hardware, with a real-time transport layer between them. |
| Conversation behaviour | running + tested | Automated invariant tests plus a scripted harness that holds simulated conversations and flags advice-giving, hearsay, repetition and invented facts. |
| Escalation tiers | specified | Fully specified including the safeguarding route and its five-year retention. Implementation in progress. |
| Morning board | sample live | A working interface is published with clearly labelled fictional data, so operators can see it before a single resident is enrolled. |
| Family view | sample live | Same, with per-person revocation and single-use enrolment codes specified. |
| Telephony | designed | Carrier plan, number pooling, caller-ID attestation, consent and calling-window model, voicemail and retry logic, dual-carrier failover. No calls are being placed today. |
| HIPAA / BAA posture | specified | Template not yet drafted. Becomes urgent at the first skilled-nursing prospect. |
| Independent security assessment | not yet | None commissioned. We will commission one before general availability rather than claim coverage we do not have. |
| Live communities | none | No community is live and no resident information is being processed by us today. |
What it has to do at seven in the morning
The hour that matters, and what happens to it when your census grows.
A check-in service is only worth having if the board is ready before your staff walk the floor. Everything below exists to make that true on the day you are short-staffed, not just on the day of the demo.
Every resident, at once
The morning's conversations happen in parallel, not one after another. Forty residents and a hundred and forty produce a board at the same time. Your round does not stretch because you grew, and nobody is waiting in a queue to be asked how they slept.
Warm before the first call
Morning rounds are the most predictable workload there is — same window, every day. Capacity is prepared against your schedule in advance, so the first resident of the day gets the same call as the fiftieth. No cold start, no slow first conversation.
Two carriers, not one
Telephony is the part most likely to fail, and the part nobody thinks about until it does. The design carries a primary and a secondary carrier with failover, because a morning where the calls do not go out is a morning your staff learn not to rely on it.
On price. We are not publishing one. Pricing is quoted per community, by a person who has looked at your census, your calling window and how your shifts actually run — and it stays unpublished until we have run a real community and know what it costs to serve one well. A number invented for a website before that would be a number we would have to walk back in front of you.
What GRACE may not do
Limits written down so they can be held against us.
- It does not give advice. Not medical, not financial, not "have you tried". The moment it starts advising it becomes an unlicensed practitioner with a telephone.
- It does not embarrass people who forget. No "you told me last week", no quizzing, no correcting. For someone with dementia, being called out by a machine is a small cruelty that compounds. Memory is used to be kind, or not at all.
- It does not pretend to be human. It says what it is. It does not invent a weekend, a family, or a bad day at work to seem relatable.
- It does not diagnose. It reports what was said and what changed. A clinician decides what that means.
- It does not investigate. When a resident discloses harm, it acknowledges, it stops, and it routes. Interrogation contaminates an account a real investigator will need.
- It does not put safeguarding on the shared board. Not once, not by accident, not for convenience.
- It does not keep everything. Retention is bounded and published, with one deliberate exception for safeguarding records.
One turn of one call
The same engine. Everything up to the check is identical code.
Hearing
Speech recognition tuned for the telephone band rather than a studio microphone, running on GPU so the transcript arrives while the sentence is still warm. Silence and noise get misrecognised as words by every system on the market; ours detects that pattern and refuses to act on it rather than answering a question nobody asked.
Thinking
A language model does the talking, but it does not decide what the turn is about. A priority ladder does — repair first, then an outstanding request, then humour, then following up on something the person just said. The model is handed one job at a time. Models drift when you ask them to hold six goals at once; this is the mitigation.
Speaking
Voice synthesis with one consistent voice, because a companion whose voice changes between calls is not a companion. Generated audio is cleaned before it is sent — synthesis models emit a burst of high-frequency noise at clip onset, which is inaudible on a laptop and unpleasant in a hearing aid.
Repairing
When the system does not understand, it does not say "I didn't catch that" three times. A three-rung ladder makes the question lighter, then easier, then takes the blame itself and changes the subject. Nobody should have to feel tested by a telephone.
The part not allowed to be clever
The same check, and a completely different idea of who is entitled to hear it.
A person at home has no staff, no shift and no administrator. Routing their conversation to anyone would betray the only reason they called a service where nobody takes notes. So kAI escalates to the person themselves — and to exactly one other place, once.
| Situation | What kAI does |
|---|---|
| Danger | Asks them to hang up and call 911. Nothing follows that line — it does not keep talking, and it does not try to handle it itself. |
| The line then goes quiet | Only then, and only if no one responds, does it alert their emergency contact — by every means available. |
| Anything else | Nothing. It listens, and the conversation ends when they end it. |
It never claims a contact has been alerted unless one truly has. A system that says "I've let your daughter know" when it hasn't is worse than one that says nothing, because the person stops looking for help.
What does not exist on this side: no staff board, no dashboard, no family view, no tiers, no administrator, no summary written for anyone to read. Nothing from a kAI conversation can appear in any kAIndness LLC system — not a transcript, not a flag, not a count, not whether the phone was answered.
What is running, and what is not
The free service is earlier than the paid one. Saying so is the point.
| Component | State | Detail |
|---|---|---|
| Voice stack, end to end | running | The same pipeline, on the same codebase. This is what a royalty-free licence actually buys the charity: a finished engine it did not have to fund. |
| Conversation behaviour | running + tested | The companionship harness holds simulated conversations across scenarios — a hectic day, grandchildren, jokes, mishearing, a lonely evening — and flags advice-giving, hearsay, repetition and invented facts. |
| Memory kindness | running + tested | Invariant tests assert kAI never quizzes, corrects or references a past conversation unprompted. These run on every behaviour change. |
| 911-first escalation | specified | The rule, the two-minute silence trigger and the contact notification are specified. Implementation in progress. |
| Nonprofit server | not yet | Not stood up. It runs in The kAIndness Project's own cloud accounts, funded by the nonprofit, and cannot be built on the LLC's. |
| Telephony | designed | One consistent number per person, caller-ID attestation, calling windows and a DTMF opt-out. No calls are being placed today. |
| Nonprofit status | in formation | The kAIndness Project has not received a determination letter. Nobody can be enrolled until it exists properly. |
Given away, so it has to cost almost nothing
A free service dies of its own infrastructure bill before it dies of anything else.
This is the constraint that shapes the whole architecture, and it came from the charity side rather than the commercial one. A paid product can absorb an inefficient design by charging for it. A free one cannot.
Under a dollar a day
Scheduler and database are serverless; compute is requested when a call starts and released when it ends. Between calls the service costs approximately nothing to exist.
Its own accounts
The nonprofit pays its own vendors from its own cloud accounts, on its own grants. No money passes from the LLC, in either direction. The licence is the whole subsidy.
Telephony is the real bill
Carrier minutes, not GPUs, are what will limit how many people this can reach. Which is why the carrier layer is specified in as much detail as the models are.
What kAI may not do
Limits written down so they can be held against us.
- It does not give advice. Not medical, not financial, not "have you tried". The moment it starts advising it becomes an unlicensed practitioner with a telephone.
- It does not embarrass people who forget. No "you told me last week", no quizzing, no correcting. For someone with dementia, being called out by a machine is a small cruelty that compounds. Memory is used to be kind, or not at all.
- It does not pretend to be human. It says what it is. It does not invent a weekend, a family, or a bad day at work to seem relatable.
- It does not diagnose, and it does not report. Not to a clinician, not to a family member, not to us. There is no recipient. What it hears stays in the call.
- It does not build a record. No staff board, no family dashboard, no referral back to the company — nothing that could later be read by someone the person did not choose.
- It does not feed the paid product. Not a summary, not a flag, not a usage count, not whether someone picked up.
- It does not promise help it has not sent. It never says a contact has been alerted unless one truly has.