SafeShift

Detecting and Rewarding Healthy Work Environments

Comment

Inspiration

The work environment is full of invisible stressors, and employers are not taking sufficient action. Workplace ethics are often treated as promises instead of measurable responsibilities, and there is not much incentive for employers to go above and beyond to optimize their employees' working environment.

What it does

We built SafeShift, a real-time, AI-powered system that measures workplace environmental conditions using privacy-safe sensors, turns those signals into actionable risk insights, and rewards companies that maintain healthy, ethical working environments using cryptocurrency incentives. It detects various health signals such as the temperature, atmospheric pressure, magnetic field strength, and noise pollution in the workplace - all factors that can harm the health of workers if not at optimal levels. We further measure workplace morale through detection of vulgar/rude language between employees/employer, while also tracking positive feedback as well. This is all combined into a risk score, and if a certain threshold of "employee friendliness and safety" is met in the workplace, the company will receive SOL as a reward - the goal was to reward companies that treat their employees right. Blockchain was a key portion of this project, promoting publicly-visible transactions between governments/donors and companies. Companies want their reputations to be good, and having a healthy work environment for employees is a key factor. Having their SOL earnings public will prove that they are doing this.

How we built it

0. Core Features:

Solana

SafeShift uses Solana for automated reward distribution, sending SOL from a main reward wallet to company wallets when safety scores are high (risk < 0.3). Rewards range from 0.1 to 0.2 SOL, calculated dynamically based on how safe conditions are, with lower risk scores earning higher rewards. Each transaction is executed as a native SOL transfer using Solana's System Program, confirmed in ~400ms at ~$0.00025 per transaction, creating an immutable on-chain record. Transaction signatures are stored in Snowflake for cross-reference, enabling public verification via Solana Explorer for complete transparency. The system includes a 1-hour cooldown mechanism per company to prevent gaming and ensure sustained compliance, while supporting multi-tenant operations where one main wallet can distribute rewards to many company wallets simultaneously.

Why we used Solana:
  • Transparency -> any reward is verifiable, avoiding faking payouts - It's not just the sender and receiver seeing the transaction, because with Solana, other parties such as manufacturers (example: H&M wanting their subsidaries to be rewarded and ensure compensation is given). We wanted to promote the idea of improving workplace conditions. We want people to see if a company is being rewarded, because it increases a company's credibility when you know they're treating their employees right.
  • Automatic payouts -> instant reward system for companies that are doing good
  • Better alternative than money -> harder to enforce cooldowns or thresholds (such as political powers interfering with this ethical transaction of money for the greater good). Ethical behaviour becomes enforceable, not just optional this way.
  • Not just basic transfers, but policy-driven payouts, with eligibility thresholds (if a workplace meets a certain health/morale standard)
  • One main sponsor wallet x many company wallets. The sponsor wallet represents a donor, or the government, or anybody wanting to encourage companies to treat workers right
  • Storing tx signatures in Snowflake and then link payouts to actual ML decisions by the model we trained
Solana x Snowflake

Snowflake stores all sensor data, ML predictions, blockchain records, and governance metadata across multiple schemas (RAW, STAGING, ANALYTICS, ML, BLOCKCHAIN, GOVERNANCE) for comprehensive data management. The RAW schema captures unprocessed Arduino sensor readings with full JSON payloads, while the ML schema stores risk scores and confidence levels from the LSTM autoencoder model for performance tracking. Snowflake Streams and Tasks automate ETL pipelines, transforming raw data through staging layers into analytics-ready aggregations that calculate factory performance metrics and compliance scores. The BLOCKCHAIN schema maintains reward payout records with Solana transaction hashes, enabling cross-verification between on-chain transactions and internal records for audit purposes. The most advanced feature is Snowflake Cortex AI integration, which performs end-of-session analysis by querying historical sensor data, identifying safety threshold violations, and using the COMPLETE function with mistral-large2 to generate AI-powered health warnings like "Temperature was very high, can lead to heart disease" with actionable recommendations for factory managers.

1. Hardware Layer:

  • Arduino Nano 33 BLE Sense devices with built-in sensors (temperature, humidity, pressure, light, noise, magnetic field)

  • Sensors send data to the backend server every second via HTTP

2. Backend Infrastructure:

  • Node.js/Express server receives sensor data, validates it, and stores it in Snowflake
  • Python FastAPI server hosts two ML models: LSTM Autoencoder (PyTorch) for environmental anomaly detection, toxic-bert (HuggingFace Transformers) for verbal abuse classification, Solana integration using @solana/web3.js for automated reward distribution

3. Data Pipeline:

  • Snowflake data warehouse with schemas for RAW, STAGING, ANALYTICS, ML, BLOCKCHAIN, and GOVERNANCE, Automated ETL using Snowflake Streams and Tasks, Snowflake Cortex AI for end-of-session analysis and health warnings

4. Frontend:

  • React + TypeScript dashboard built with Vite
  • Real-time sensor monitoring with live updates
  • Live microphone input using Web Speech API for demo purposes
  • NLP warnings display showing abuse detection results
  • Reward panel showing Solana transaction history

5. Integration Flow:

Arduino → HTTP POST to Express server
Express → FastAPI ML endpoint for risk scoring
Express → Solana blockchain for reward transactions
Express → Snowflake for data storage
Frontend → REST APIs for real-time dashboard updates

Challenges we ran into

Solana integration complexity: Learning Solana’s transaction model, account management, and network interactions. We chose TypeScript/JavaScript over Rust to move faster during the hackathon.

Frontend/backend separation: Managing CORS, API proxying with Vite, and ensuring the frontend could communicate with both the Express server and Solana network.

ML model integration: Bridging Python FastAPI with Node.js Express, handling model loading times (toxic-bert is ~500MB), and implementing fallback scoring when the ML API is unavailable.

Real-time data flow: Ensuring sensor data flows from Arduino → server → ML → blockchain → frontend with proper error handling and state management.

Privacy concerns: Designing the NLP system to not store transcribed text (privacy by design) while still providing useful abuse detection metadata.

Accomplishments that we're proud of

End-to-end system: From hardware sensors to blockchain rewards to frontend dashboard, all components working together.

Automated reward system: Real-time risk assessment triggering automatic Solana transactions when safety scores are high.

Dual ML approach: Combining environmental anomaly detection (LSTM) with verbal abuse detection (NLP) for comprehensive safety monitoring.

Privacy-first design: NLP system analyzes text without storing it, protecting worker privacy while enabling abuse detection.

Real-time dashboard: Live sensor monitoring, instant NLP feedback, and historical warning display all in one interface.

AI-powered insights: Using Snowflake Cortex AI to generate actionable health warnings from historical data patterns.

What we learned

Blockchain integration: Solana’s transaction model, account management, and the importance of proper error handling for blockchain operations.

ML model deployment: Balancing model accuracy with inference speed, handling large model files (toxic-bert), and implementing graceful fallbacks.

Real-time systems: Managing state across multiple services (Arduino, Express, FastAPI, Snowflake, Solana, React) and ensuring data consistency.

Privacy engineering: Designing systems that provide value while protecting sensitive data (no text storage in NLP pipeline).

Data warehouse design: Structuring Snowflake schemas for scalability, using Streams and Tasks for automated ETL, and leveraging Cortex AI for advanced analytics.

What's next for SafeShift

Enhanced hardware integration: Deploying Arduino devices in real factories, improving sensor calibration, and adding more sensor types (CO2, PM2.5, VOCs).

Advanced ML models: Training on real factory data, improving anomaly detection accuracy, and adding predictive capabilities (forecasting unsafe conditions before they occur).

Expanded reward mechanisms: Implementing SPL tokens for custom reward currencies, time-based vesting schedules, and insurance discount calculations.

Built With

Share this project:

Updates