Back to satwik's profile
11views

WardZero — Care with continuity

Explained home-care concern that survives sensor outages and tracks a caregiver's response.

satwik sharma
  • Video of it working
  • Problem, approach and result explained
  • Public source code
  • Commits by the candidate· 29 of 29 commits
  • Built during the challenge
  • Answered the full brief· 6 of 6
  • Expert reviewed
WardZero — Care with continuity

3/3

Queued readings recovered in demo

73

Automated behavior checks passed

8

Chrome journeys passed

Walkthrough

Overview

The problem

A home-care handoff can fail even when a dashboard displays every reading. A caregiver still needs to know what changed, why it matters, who owns the concern, and whether anyone responded. Connectivity can fail while an unanswered alert sits unnoticed.

WardZero explores this gap with one synthetic post-discharge patient and two care roles. It connects simulated readings and English caregiver notes to explained concern, an assigned alert, and a recorded human response. Missing or old evidence stays visible. The prototype does not diagnose patients or claim validated 24–48-hour prediction.

The approach

I built WardZero in Next.js and TypeScript with filesystem-backed embedded PostgreSQL. A browser simulator saves readings in IndexedDB before sending them through a validated API. Stable event identities prevent duplicates during retries, while observation time remains separate from receipt time.

Gemini extracts allowlisted observations from English caregiver notes, preserving exact excerpts, negation, and historical/current context. Application rules combine validated observations with fresh vitals to explain Low, Medium, or High concern. Insufficient or stale evidence produces Unknown. The model cannot execute care actions.

Each active alert has an owner. Acknowledgement, recorded action, and resolution are separate persisted steps. A server worker escalates unanswered alerts to a backup coordinator, even with the browser closed. Isolated runs and visibly accelerated demo time make scenarios repeatable.

I separated patient concern, data freshness, and response state so that resolving an alert does not rewrite concern, and replaying old readings does not create a fresh emergency.

The result

The working care loop shows stable readings establish Low concern, live Gemini observations raise it to Medium, and deteriorating vitals raise the same alert episode to High. The assigned caregiver acknowledges, records an action, and resolves with a reason; the audit trail remains available.

In the recorded outage scenario, three queued readings survive refresh. Reconnection restores all three, producing six stored readings and an empty queue. Separate checks cover failed requests and lost acknowledgements without duplicate storage.

The verified improvement release passes 73 automated tests and eight Chrome journeys. Its final 40-case synthetic live Gemini evaluation matches all expected cases; earlier imperfect runs are retained. These are small development examples, not clinical accuracy evidence.

A controlled storage-backed replay raises concern 45 synthetic seconds before its authored individual-threshold marker, versus 40 seconds with trends alone. Inference wall time is excluded. Negative, historical, stable, and failed-analysis controls make the comparison explicit. Production-worker checks verify browser-closed escalation, while local HTTP tests verify optional email retry and recovery. Source, setup instructions, and verification evidence are available in the public repository.

Reflection

The main lesson was that recovery and ownership belong in the data model, not only in dashboard labels. Tests and fresh review exposed subtle problems with future timestamps, mixed negation, historical context across clauses, and notification retry behavior. Failing regressions reproduced these issues before the fixes were accepted.

WardZero remains a local, single-patient prototype using synthetic identities, readings, and notes. Its engineering thresholds and small note evaluation are not clinical validation. The early-warning timeline is constructed and does not establish real-world lead time or 24–48-hour prediction. Human caregiver usability testing has not been performed.

In-app escalation is enabled. Optional email is implemented but disabled by default; only local HTTP and persistence tests have been performed, so real cloud or mailbox delivery is unverified. State and overdue deadlines recover after restart, but monitoring cannot continue while the server or computer is off. A real deployment would require production authentication, hosted durable storage, a supported worker, clinical governance, and broader independent evaluation.

Links & files

Artifacts

2

Gallery

5