AI Enterprise SolutionsGPT & Claude

Algorhythm

THE TARGET USER: Seniors and vulnerable patients living independently who are at high risk of falls and cardiovascular anomalies, alongside the family caregivers, home-care nurses, and clinicians responsible for their safety. Caregivers currently face constant anxiety and severe alert fatigue, while elderly individuals want to preserve their autonomy and privacy without invasive surveillance cameras or clunky medical pendants. THE PROBLEM: Falls are the leading cause of fatal injury among seniors, and over 80% of victims cannot press a manual emergency button due to shock, loss of consciousness, or physical injury. Lying unattended on the floor for hours leads to rapid clinical deterioration, hypothermia, and high mortality. Existing solutions fail in the real world: manual panic buttons require the victim to be conscious; smartwatches trigger frequent false alarms from everyday arm gestures; and video cameras are an unacceptable invasion of privacy in bedrooms and bathrooms where most falls occur. Furthermore, current monitoring systems spam caregivers with raw, isolated numbers without clinical context, making it impossible to distinguish between normal activity and life-threatening emergencies.

GitHub

Preliminary score

52.0/ 100Earlyaverage of 2 judges

Value for the user17.0 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI13.5 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing11.0 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility5.5 / 15
Data requirements, running costs and a clear next step.
Originality5.0 / 10
A meaningful difference from common solutions.

What the judges wrote

Claude Opus 5.5Preliminary48/100
Strengths
Your problem framing is concrete: seniors living alone who cannot press a panic button after a fall, and caregivers swamped by context-free alerts. The three-tier escalation (low, medium, high risk) gives a clear outcome model. Your prototype has real engineering behind it: an Expo/React Native app, an Express and WebSocket gateway, and a scenario simulator covering 10 clinical and device-failure cases, including stale data, offline radar and failed notification delivery. Reporting specific bugs you hit, such as the rule-precedence collision and the multi-patient telemetry leakage, is honest and useful.
To improve
Your AI contribution is unclear. The README describes a deterministic rule engine, and the GPT-4o-mini triage and summary layer is claimed but not shown, so demonstrate one LLM-produced incident summary or escalation decision and compare it with the rules alone. Your tests confirm that the rules behave as coded, but not that fall detection is accurate. Even a small recorded ESP32 or IMU dataset with false-positive and false-negative counts would show whether you actually reduce alert fatigue. Your feasibility case needs hardware sourcing, LLM and connectivity running costs under a no-subscription 199 AZN model, medical-device and regulatory considerations, and a realistic pilot partner such as a home-care agency.
GPT-6 AstraPreliminary56/100
Strengths
Your project addresses a concrete caregiving problem: detecting emergencies without cameras or relying on someone pressing a button. Your documented simulator connects telemetry scenarios to risk levels, incident evidence, and caregiver acknowledgement, including device failures rather than only happy paths. Your testing report names specific bugs and fixes, especially stale-data rule precedence and cross-patient telemetry leakage.
To improve
Your README documents a deterministic workflow; show the LLM executing on a scenario and measure its contribution beyond rules. Extend your tests to held-out sensor recordings, measuring missed falls, false alarms, and alert latency against a wearable-only baseline. Before automatic emergency escalation, define clinician-reviewed thresholds, confirmation safeguards, and delivery-failure handling. Specify the actual radar hardware, labeled-data requirements, per-user operating costs, and a supervised home-care pilot; the 199 AZN selling price does not establish running costs.

The team

Nurgun Hajırahimova
product
Yaghmur Mammadov
developer
Ulker Soltanova
product
Elnur DautovLead
business
Eyub Azizov
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: Hybrid Risk Assessment Architecture: Deterministic clinical rule validation paired with an LLM-based clinical decision layer (GPT-4o-mini / medical LLM prompt pipeline). The LLM processes combined multi-modal device telemetry (bracelet vitals + ESP32 radar/IMU data), patient medical baseline context, and post-incident stillness duration to perform real-time triage reasoning, contextual escalation decisions, and natural language clinical incident summaries for caregivers.
Data: Synthetic physiological vitals (optical PPG heart rate, SpO2, skin temperature, blood pressure) and simulated ESP32 60GHz mmWave radar & IMU movement telemetry calibrated to clinical safety thresholds, enriched with clinical prompt scenarios for LLM evaluation.
Components: React Native, Expo SDK 57, Expo Router, TypeScript, Zustand, React Native Web, Express.js & WebSocket gateway, Vitest, Zod, LLM Integration Interface (OpenAI / Anthropic SDK).
Built during the event: Confirmed

Quality testing

What we tested: We wrote 11 automated unit tests in Vitest testing our pure deterministic risk assessment engine across 10 clinical and hardware failure edge cases: baseline monitoring, minor arrhythmia, fall with recovery, fall with prolonged inactivity, hypoxia/elevated temperature, bracelet disconnection, ambient ESP32 radar offline, stale data (>10m), notification dispatch failure, and multi-sensor fusion (fall + inactivity + abnormal vitals), plus deduplication window logic. What broke during development: 1. Rule Precedence Collision: Stale data (>10m valid data) was initially matching the "Bracelet Disconnected" rule instead of "Stale Measurements" because both evaluated validity status. We resolved this by strictly ordering rule evaluation and separating device presence from timestamp staleness. 2. React Key Collisions in High-Frequency Telemetry: Rapid scenario simulation caused identical mock data IDs to collide in the time-series bar chart, triggering reconciliation warnings. We fixed this by introducing composite temporal keys (${id}-${measuredAt}-${index}). 3. Multi-Patient Telemetry Leakage: Child components originally read un-scoped arrays from the global Zustand store rather than filtering by personId, causing potential telemetry overlap. We introduced strict patient-scoped selectors. Comparison with today's solutions: Traditional emergency pendants require conscious activation (failing in >80% of falls with head trauma or shock). Smartwatches suffer high false-positive alarms from hand gestures. SmartCare solves this through dual-sensor fusion: correlating spatial room radar (ESP32) with wearable wrist telemetry (HR spikes, oxygen saturation drops, and post-fall motionless duration) to verify genuine emergencies before escalating alerts.

Worth 20 of the 100 points.

Running it

Prerequisites: Node.js 18+ (tested on Node v20/v24). No hardware required (fully simulated telemetry & dev tools included). 1. Clone and run the Frontend / Mobile Web App: cd mobile npm install --legacy-peer-deps npm run web -> Opens at http://localhost:8081 in any browser. 2. Run Automated Test Suite (11/11 passing tests): cd mobile npm run test npm run typecheck 3. Run Backend Ingestion Gateway (Optional): cd backend npm install npm run dev -> Runs REST API and WebSocket stream at http://localhost:4000. How to test the live app: - On the dashboard, tap the "Simulate" button to launch the scenario runner. - Select any of the 10 clinical scenarios (e.g., "Fall + Prolonged Inactivity", "Abnormal Vitals", "Bracelet Disconnected") to watch the multi-sensor fusion engine evaluate risk and escalate alerts in real-time.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.