๐Ÿซ‚

Safety Kaki

Private demo โ€” please enter the access password.

Safety Kakiๅฟƒ่ทฏ ยท keeping our family safe, together
v1 visual prototype โ€” navigable mock-up, not for clinical use. No real data or backend.
9:41๐Ÿ“ถ ๐Ÿ”‹
๐ŸฆŠ

Hi, Sarah's family ๐Ÿ‘‹

Your Safety Kaki is here to help everyone stay safe & connected

๐Ÿ‘จโ€๐Ÿ‘ฉโ€๐Ÿ‘ง The Tan family ยท Kaki since Jun 2026

Sarah's Kaki Band

Worn today ๐ŸŒฟ
๐Ÿ”‹ 82% ยท ~14h left
BANDโœ“ working well
COVERED TODAY18/24h
LAST SYNC2 min
โ„น๏ธ No alerts so far today โ€” and the band wasn't recording for about 6 hours (school, charging, bath time). That tells you what it noticed, not that everything is okay. If something happened, you can always tell Ms Lena.
๐Ÿ›ก๏ธ Our Safety System

This is the plan we made together with our Protection Officer, Ms Lena. If things get heated at home, this is what happens โ€” step by step. Everyone knows their part. ๐Ÿ’›

0

A gentle nudge, first

If Dad's stress rises, his phone quietly suggests his own calm-down step โ€” going to the balcony for a bit (not breathing, which doesn't help him).

๐ŸŒฌ๏ธ For: Dad ยท self-regulation
1

Mum gets a message

If it keeps rising, Mum gets a WhatsApp to come and help settle things. She taps โ€œI've got thisโ€ so everyone knows she's on the way.

๐Ÿ’š Safe adult: Mum (lives here)
2

Grandpa is next

If Mum can't respond in a few minutes, Grandpa (5 mins away) is messaged next. He can see if Mum's already handling it, so nobody over-worries.

๐Ÿ’™ Safe adult: Grandpa Tan
3

Ms Lena is looped in

Our officer is told what happened. For our family, she does a friendly follow-up within a day โ€” she doesn't need to call in the moment unless it's serious.

๐Ÿ“‹ Protection Officer ยท follow-up in 24h
4

If none of us can be reached

If nobody answers and Sarah might not be safe, the Protective Service comes to check in person. We talked about this step together when we made the plan โ€” it isn't a punishment, it's the last safety net, and it's here so Sarah is never left waiting.

๐Ÿ›Ÿ Protective Service ยท last resort ยท always available

โœ๏ธ It's your plan โ€” you can change it

Anything here can be talked through and edited with Ms Lena. You can also export a copy or pause the plan when the family steps down to low-risk.

๐Ÿ”ต Circles of Safety & Support

The people around our family who can help. The closest circle already knows everything and are part of our safety plan.

Wider circle
Knows a little
Inner ยท knows all
๐Ÿ‘ง Sarah
๐Ÿ’šMum
๐Ÿ’™Grandpa
๐Ÿง•Aunt Mei
๐Ÿ‘ฉโ€๐ŸซMs Lena
๐ŸซTeacher
๐Ÿ˜๏ธNeighbour

๐Ÿ’š = active safe adult in our plan

๐Ÿก Sarah's Safety House

Sarah drew this with Ms Lena. It shows the rules that help her feel safe, and where she feels she is on the path to safety right now.

๐Ÿก
โ€œMy safe homeโ€
๐Ÿ  House rules that keep me safe
โœ… Grown-ups use calm voices
โœ… If it gets loud, I can go to Grandpa's
โœ… Mum always answers her phone
โœ… No scary surprises
My path to safety
๐Ÿ‘ง

Sarah places herself near the safe end ๐Ÿก โ€” โ€œmostly safe, getting better.โ€

๐Ÿ’ช Our Family Strengths

Things that are already going well for us. Ms Lena adds new ones she notices, and โญ marks the ones we've celebrated together.

Dad went to all his counselling sessions Mum is calm under pressure Close bond with Grandpa Sarah loves school Family eats dinner together Dad apologises & repairs Neighbours look out for us
๐Ÿ“š Resources for My Family

A starter set, plus anything Ms Lena adds just for us.

๐ŸŽฎ

Calm-Together mini games

2-player breathing & feelings games for Sarah + a parent

โ€บ
โ˜Ž๏ธ

NAVH Helpline ยท 1800-777-0000

24/7 support, any time you need to talk

โ€บ
๐Ÿงก

Positive parenting tips

Short, practical, Singapore-context ideas

โ€บ
๐Ÿซ

Fei Yue Family Service Centre

Added by Ms Lena ยท programmes near you

โ€บ
๐Ÿ‘ง

Sarah Tan

9 ยท returned home 3 wks ago ยท moderate-to-low risk ยท PO: you ยท sample data

๐ŸŸข Band worn ยท 82% ๐Ÿ”” Alerts: 1 this week ๐Ÿ“ˆ Safety 7/10 โ†‘

โš ๏ธ What are we worried about?

Harm ยท complicating factors
  • One incident of physical harm by father while under the influence of alcohol (Jun 2026).
  • Complicating: work stress, past alcohol use, thin weekday support.

โœ… What's working well?

Protection & belonging ยท strengths
  • Father remorseful, first-time, completed counselling.
  • Mother + grandfather active, reliable safe adults nearby.
  • Sarah attends school, strong bond with grandfather.
๐Ÿ’ช Family strengths ยท โญ = shared with family
Father completed counselling Mother calm under pressure Strong grandfather bond Sarah engaged at school Repairs after conflict Neighbour support

๐Ÿ”” 2 strengths not yet shared โ€” raise at next visit.

๐Ÿ“ Safety & Wellbeing Scale

0 = not safe enough to be home ยท 10 = safe enough to close ยท different people's views
Sarah 6
PO 7
Mum 8
0 ยท danger10 ยท safe to close

๐Ÿ“ˆ Safety scale over time

Child indicates at check-ins โ†’ you log it here. Kept off the family app so parents aren't confronted by the trend.
4
Wk1
5
Wk2
5
Wk3
6
Wk4
7
Wk5

๐Ÿก Safety House progress

Child-voice tool completion
  • โœ… House drawn with Sarah
  • โœ… Safety rules agreed
  • โœ… Safe adults confirmed
  • โฌœ Review with network โ€” due wk 6
๐Ÿ”ต Circles โ†’ safe adults (exportable to case file)
๐Ÿ’š Mdm Tan (Mum) ยท 9xxx1234 ๐Ÿ’™ Mr Tan (Grandpa) ยท 9xxx5678
โฌ‡๏ธ Export contacts to case recording

๐ŸŽฏ What this system is for โ€” and the rules it runs under

A complement to your intervention, never a replacement for it.
Purpose
  • Deterrence โ€” the household knows monitoring is active and agreed. Behaviour that changes because someone is accountable is a legitimate outcome, not a missed detection.
  • Accountability โ€” an agreed safety plan with observable follow-through, reviewed openly with the family.
  • Mobilising support โ€” bringing the family's own network in early, so response never depends on the child reporting.
Data governance ยท PDPA
  • No raw audio ever leaves the band โ€” acoustic features and event labels only. Speech is not recorded, transmitted or stored.
  • Collection limited to purpose โ€” physiological and acoustic-event data for safety monitoring only. No location tracking, no content capture.
  • Defined retention and deletion, family-visible thresholds, and an access log of who viewed what, when.
  • Operates under PSV's statutory safeguarding role, with PDPA governing how data is collected, used, retained and disposed of.

๐Ÿ”” Alert ladder for this family

You configure the response per case โ€” reliable network here means Tier 3 is a same-day follow-up, not a live video call. Co-designed in the safety-planning session.
0

Self-regulation nudge โ†’ Dad's app

Custom: โ€œstep out to the balconyโ€ (breathing didn't suit him). Affirms when he follows the plan.

1

Safe adult ยท Mum (in-home)

WhatsApp ยท must acknowledge within 3 min, else escalate

2

Safe adult ยท Grandpa (5 min away)

Notified next ยท sees if Mum already responded

3

Protection Officer (you)

Follow-up within 24h ยท not live call for this case

4

PSV intake ยท unannounced safety check

Auto-fires on silence, tamper-with-distress, or an implicated safe adult โ€” not left to PO discretion

๐Ÿ›‘ When the safe adult is unavailable โ€” or is the risk

The ladder never assumes the in-home adult is safe. These rules override the tier order and cannot be switched off by the family.
  • Person-of-concern flag โ€” any adult in the plan can be marked by you as a person of concern. They are never a Tier 1/2 recipient, and are not told that an alert fired.
  • Implicated-adult bypass โ€” if the signal itself implicates the in-home adult, Tier 1 is skipped: the alert goes straight to the external safe adult and to you.
  • Silence is escalation โ€” no acknowledgement from any safe adult within the agreed window auto-escalates to Tier 4. Silence is never read as โ€œresolvedโ€.
  • Tamper + distress โ€” a band removed or offline while a distress signal is active is treated as escalation, not as a device fault.
  • Child-initiated โ€” the child can call for help directly from the band, bypassing every tier.

Escalation to PSV โ€” including removal โ€” stays available at every stage. The system never closes that door.

๐ŸŽš๏ธ Detection sensitivity

Tunes how readily a signal escalates โ€” not whether it is noticed. Thresholds are personalised to this child's own baseline, not fixed population cut-offs.
Low
Balanced
High

Thresholds are transparent and reviewed with the family โ€” no hidden surveillance.

๐Ÿ“ Baseline state โ€” sensitivity means nothing without it
โ—

Sarah โ€” MATURE

Coverage reached across required contexts: rest, sleep, weekday, weekend. Maturity is earned by data coverage, never by a calendar timer.

active
โ—

Marcus โ€” PROVISIONAL

Some personal reference exists, context coverage incomplete. Alerts carry a reduced-confidence flag.

day 9
โ—

Wei โ€” REQUALIFICATION REQUIRED

Strap and wrist changed โ€” earlier data no longer comparable.

paused

โš ๏ธ A baseline records what we observed during calibration โ€” not the child's healthy or safe state. If a child was already living under chronic stress while we calibrated, that becomes their โ€œnormalโ€ and later harm may raise no anomaly; conversely, genuine improvement can read as deviation. This limit is clinical, not technical.

๐Ÿง  Signal fusion โ€” agreement raises the tier, it never gates the alert

Requiring two signals to agree would cut false positives by manufacturing false negatives in exactly the quietest cases. So concordance escalates; it is never a precondition for being noticed.
โœ“โœ“

Both agree โ€” physiological + acoustic

Highest confidence โ†’ immediate Tier 1 alert to the safe adult.

โœ“

Physiological only โ€” stress rise, no sound

Never discarded. Logged, pattern accumulated, surfaced at your next check-in. This is the path that covers quiet threats and closed doors.

โœ“

Acoustic only โ€” raised voices, no arousal

Logged the same way. The child may be absent, asleep, or shut down.

?

Neither โ€” or coverage degraded

Recorded as uncertainty, not absence. โ€œNo alertโ€ is never rendered as โ€œsafeโ€.

Fusion requires aligned timestamps, in-range sensors and genuinely independent outputs. Stale, masked or out-of-range streams degrade coverage rather than silently passing as normal.

โฑ๏ธ Freshness gate โ€” a live alert needs two conditions, not one

Every record carries observed_at, received_at and a clock-uncertainty estimate. An urgent alert fires only on anomaly detected AND received within 120 s AND clock uncertainty within tolerance. A reading that arrives three hours late is still valuable evidence for review โ€” but it is never presented as live. A dashboard showing a recent-looking heart rate that was actually recorded hours ago is a safety hazard, not a display bug.

๐Ÿ”— Sensor dependency โ€” three numbers can be one sensor

Heart rate, HRV and any vendor โ€œstress scoreโ€ are commonly derived from the same PPG sensor โ€” they are not three independent confirmations, and a contact failure takes out all three at once. Likewise shouting and crying may come from one acoustic model, and steps and sleep from one accelerometer. The system holds an explicit dependency graph so agreement between two outputs of the same sensor is never counted as corroboration.

๐Ÿ‘๏ธ What this system does not see

Stated plainly so monitoring data is never mistaken for full coverage. Environmental sensing corroborates โ€” it does not close the observation gap.
  • Quiet threats and coercive control โ€” harm without volume
  • Incidents behind closed doors, or in a different room from the sensor
  • Physical acts without a loud impact
  • Anything occurring outside the home
  • Periods when the band is off-wrist or out of range, or audio is masked by television or other children
  • Harm that produces no arousal at all โ€” freezing, withdrawal, dissociation and learned compliance are common child responses to threat
  • Much of what happens during struggle or vigorous distress โ€” wrist sensors degrade under exactly the high-motion conditions a serious incident produces. Quiet periods give the cleanest data; escalating ones give the worst

This system lowers the reporting burden on the child and shortens response time. It does not replace your judgement, home visits, or the child's own account.

Mapped directly to MSFโ€™s four stated Metrics of Success โ€” PO interventions, child safety scale (0โ€“10), recurrence of abuse, and out-of-home placements โ€” with response-time evidence alongside.
๐Ÿš— PO interventions
โˆ’38%
โ–ผ de-escalated before PO needed
๐Ÿ˜Š Child safety scale (0โ€“10)
4 โ†’ 7
โ–ฒ at PO check-ins
๐Ÿ” Abuse recurrence
0
โ€” since return
๐Ÿ  Out-of-home placement
Avoided
โ–ฒ confident preservation call
โšก Time to first response
3 min
โ–ฒ vs next scheduled visit
๐Ÿ’› Relationship (smiley)
๐Ÿ™‚โ†’๐Ÿ˜„
โ–ฒ โ€œbetterโ€

๐ŸŽฏ Detection accuracy & reliability

MSF requires low false-positive and false-negative rates, with transparent thresholds and validation. Both are tracked here; every alert is reviewed with the family and sensitivity re-tuned.
๐Ÿ”” Alerts triggered
3
โ€” this month
๐Ÿ’š Resolved by network
2
โ–ฒ before PO needed
๐Ÿ“‹ Escalated to PO
1
โ€” per agreed protocol
๐ŸŽš๏ธ False positives
1
33% ยท sensitivity re-tuned
๐Ÿ” False negatives
Not verifiable
โ€” see limits below
โšก Detection latency
6 s
โ–ฒ observed โ†’ received

We deliberately do not report a false-negative count. A child may freeze, withdraw or dissociate rather than show arousal, so an unsafe interaction can occur with no physiological signal at all. Absence of an alert is not evidence that nothing happened, and any figure we printed here would be unverifiable. This is a limit of the method, not a tuning problem.

๐Ÿ“‹ Check-in inputs (feed the dashboard automatically)

Captured in-session โ†’ no double documentation. Rolls up to your supervisor & PSV.
๐Ÿ˜Š

Safety scale

Sarah's felt safety at home

7/10
๐Ÿ’›

Parentโ€“child relationship

5-point smiley Likert

๐Ÿ™‚ 4/5
๐Ÿ“ˆ

Change since last visit

much worse โ†’ much better

Better
๐ŸŒž

Added to House of Good Things

โ€œDad reads to me nowโ€

+1
๐Ÿ—ฃ๏ธ

Network / school feedback

Teacher: settled, happier

โœ…
๐Ÿ“ค One-tap: file to case record & supervisor dashboard

โŒš Device fleet โ€” my cases

Device health is never a safety conclusion. Battery is shown as time remaining, not just a percentage โ€” 20% may mean hours or minutes depending on radio conditions. Ordinary causes are assumed before interference (see the failure-order panel below).

๐Ÿ‘ง Sarah

Worn ยท good signal

82% ยท ~14h left ยท synced 2 min ago

๐Ÿ‘ฆ Marcus

Worn ยท poor contact

64% ยท ~11h ยท PPG contact weak 25 min โ€” usable data below threshold

๐Ÿง’ Aisha

Not returned after charging

96% ยท charged 07:10, not re-worn since โ€” 2h 40m unobserved

๐Ÿ‘ฌ Wei

Sync delayed ยท buffering

48% ยท gateway offline 40 min ยท device recording locally, data not yet received

๐Ÿ“ถ Coverage is layered โ€” nominal wear is not coverage
1

Nominal wear โ€” 94%

Device reports being on a body.

2

Good-contact wear โ€” 81%

Optical sensor contact passes quality checks.

3

Usable-data wear โ€” 73%

Measurements survive artefact rejection.

4

Synchronised operational coverage โ€” 68%

This is the only number that means the decision system could actually have seen something. A case can show 94% wear and 68% real coverage.

๐Ÿ”Ž Missing data is explained in this order โ€” never jump to tampering
1๏ธโƒฃ

Expected gaps โ†’ then routine adherence โ†’ then technical fault

Bathing, sport, school removal, charging. Then forgetting to re-wear, loose fit, refusal. Then flat battery, lost pairing, gateway outage, storage full, clock drift. Routine causes dominate, especially early in a deployment.

Assumed first
4๏ธโƒฃ

Unexplained or recurrent โ€” pattern, not incident

Repeated nonwear at similar times, repeated gateway disconnection, recurring gaps around particular household contexts. Raised to you as a pattern for human judgement, never as an automatic conclusion.

Review
5๏ธโƒฃ

Suspected interference โ€” last, and never certain

The band is safely removable, tamper-evident and operationally observable โ€” deliberately not "tamper-proof", which would risk entrapment, swelling and blocked emergency or medical removal. Interference can be detected; intent cannot. Gateway offline at 20:10 does not distinguish a power cut from a pulled plug.

Circumstantial

๐Ÿ“š Resources shared with this family

Add resources here โ†’ reminder to share at next session, then auto-appears in the family app.
๐ŸŽฎ

Calm-Together games

Shared โœ…

โœ“
๐Ÿซ

Fei Yue FSC programmes

Shared โœ…

โœ“
๐Ÿงก

Positive parenting series

Queued โ€” share next visit

๐Ÿ””

๐Ÿ”” How a moment of risk plays out

๐Ÿ“ฑ The family's safety network