SehatRozgar — Health + Financial Resilience Intelligence for Workers

When health stops work, resilience becomes measurable.

Inspiration

For a daily-wage worker, getting sick does not only create a health problem — it can immediately create a financial problem.

A construction worker, delivery rider, factory worker, driver, or welder may depend directly on their ability to work each day. If a health event prevents them from working, two consequences can happen at the same time:

Health Event → Medical Needs

Health Event → Unable to Work → Lost Income → Household Financial Pressure

Most healthcare systems focus on the medical side, while financial tools focus on money. We wanted to connect these two consequences and make the relationship visible.

That led us to build SehatRozgar, a worker health and financial resilience intelligence platform combining empirical machine learning, deterministic financial analytics, occupational hazard context, and AI-powered explanations.

Our core idea is simple:

When a daily-wage worker gets sick, the cost isn't only healthcare. Their income can stop too.

What It Does

SehatRozgar analyzes worker resilience across several separate and transparent dimensions rather than hiding everything behind one black-box score.

The platform provides:

Health Risk Analysis — identifies statistical health-risk patterns using an empirical machine-learning model.
Financial Vulnerability Analysis — measures the worker's ability to absorb an income interruption.
Savings Runway — estimates how many days existing savings can support household expenses.
Income-at-Risk Simulation — calculates potential lost earnings when a worker cannot work.
Occupational Hazard Assessment — provides rule-based hazard context based on occupation.
Emergency Preparedness Analysis — evaluates the worker's financial buffer against unexpected events.
AI Resilience Advisor — converts technical results into understandable and actionable explanations.
Workforce Analytics — demonstrates aggregate resilience patterns using synthetic worker profiles.

For example, consider a synthetic construction worker earning PKR 1,800 per day.

If that worker becomes unable to work for 14 days:

$$ \text{Income Loss} = 1,800 \times 14 = 25,200\text{ PKR} $$

Instead of displaying only this number, SehatRozgar analyzes how that interruption interacts with the worker's savings, household expenses, medical costs, dependents, and emergency preparedness.

The result is a more complete picture of worker resilience.

How We Built It

SehatRozgar follows a modular architecture where each component has one clearly defined responsibility.

Public Health Dataset
        |
        v
Data Cleaning & Preprocessing
        |
        v
Machine Learning Models
        |
        v
Health Risk Prediction
        |
        v
Model Explainability
        |
        +---------------------------+
        |                           |
        v                           v
Financial Profile           Worker Occupation
        |                           |
        v                           v
Financial Engine        Occupational Hazard Engine
        |                           |
        +-------------+-------------+
                      |
                      v
             Structured Assessment
                      |
                      v
                Featherless AI
                      |
                      v
          Worker Resilience Advisor
                      |
                      v
             Accessible Guidance
                      |
                      v
               React Dashboard

Our core architectural principle is:

AI explains. Models predict. Mathematics calculates. Rules contextualize.

This separation is intentional.

We do not ask a general-purpose language model to calculate financial vulnerability or generate medical probabilities.

Empirical Health Modeling

For the health component, we wanted to use real empirical data rather than synthetic medical outcomes.

We used the UCI Heart Disease Dataset — processed Cleveland subset, containing 303 historical observations.

The dataset includes features such as:

Age
Sex
Chest pain type
Resting blood pressure
Cholesterol
Fasting blood sugar
Resting ECG
Maximum heart rate
Exercise-induced angina
ST depression
ST slope
Number of major vessels
Thalassemia code

Our preprocessing pipeline handles missing values and categorical features before model training.

We compared three machine-learning approaches:

Logistic Regression
Random Forest
Gradient Boosting

We used a fixed stratified 75/25 holdout split with random_state=42.

Model Comparison

Model	ROC-AUC	Precision	Recall	F1
Logistic Regression	0.944	0.833	0.857	0.845
Random Forest	0.936	0.800	0.914	0.853
Gradient Boosting	0.925	0.789	0.857	0.822

We selected Logistic Regression for the prototype because it achieved the highest ROC-AUC in our experiment while also providing an interpretable baseline.

On the 76-record held-out set, its confusion matrix was:

True Negatives:   35
False Positives:   6
False Negatives:   5
True Positives:   30

These results represent experimental performance on a small historical dataset.

They do not establish clinical validity, particularly for daily-wage or occupational worker populations.

For this reason, SehatRozgar presents its output as a statistical health-risk pattern, not a medical diagnosis.

Explainable Machine Learning

A probability alone is difficult to understand.

SehatRozgar therefore includes a model explainability layer that shows which input features had the strongest influence on an individual model output.

For example, the interface can visualize contributions from features such as:

Chest pain type
Resting blood pressure
Age
ST depression
Cholesterol

This helps users understand why the model produced a particular result rather than simply presenting a percentage.

However, an important distinction is maintained throughout the application:

Model contribution does not imply medical causality.

The explainability layer describes model behavior only.

Financial Resilience Engine

The financial component is deliberately deterministic and auditable.

An LLM is never responsible for calculating financial metrics. All financial outputs are generated using predefined mathematical formulas.

Monthly Income

$$ \boxed{\text{Monthly Income} = \text{Daily Income} \times \text{Working Days per Month}} $$

Income Loss

$$ \boxed{\text{Income Loss} = \text{Daily Income} \times \text{Days Unable to Work}} $$

Gross Event Exposure

$$ \boxed{\text{Gross Event Exposure} = \text{Income Loss} + \text{Estimated Medical Cost}} $$

Health Savings Used

$$ \boxed{\text{Health Savings Used} = \min(\text{Health Savings}, \text{Estimated Medical Cost})} $$

Net Event Exposure

$$ \boxed{\text{Net Event Exposure} = \text{Gross Event Exposure} - \text{Health Savings Used}} $$

Daily Household Burn

$$ \boxed{\text{Daily Household Burn} = \frac{\text{Monthly Household Expenses}}{30}} $$

Savings Runway

$$ \boxed{\text{Savings Runway Days} = \frac{\text{Current Savings}}{\text{Daily Household Burn}}} $$

Emergency Fund Target

$$ \boxed{\text{Emergency Fund Target} = 3 \times \text{Monthly Household Expenses}} $$

Emergency Fund Gap

$$ \boxed{\text{Emergency Fund Gap} = \max(\text{Emergency Fund Target} - \text{Current Savings}, 0)} $$

Event Coverage Ratio

$$ \boxed{\text{Event Coverage Ratio} = \frac{\text{Current Savings}}{\max(\text{Net Event Exposure}, 1)}} $$

Remaining Savings

$$ \boxed{\text{Remaining Savings} = \max(\text{Current Savings} - \text{Net Event Exposure}, 0)} $$

Because these calculations are deterministic, the same inputs always produce the same financial outputs.

Income Interruption Simulator

One of SehatRozgar's main features is the Income Interruption Simulator.

It answers a simple but important question:

What happens financially if this worker cannot work?

The system can simulate different periods of work interruption, such as:

3 days
7 days
14 days
30 days

For each scenario, SehatRozgar can show:

Lost income
Estimated medical expenses
Gross financial exposure
Net financial exposure
Remaining savings
Savings runway
Emergency-fund gap
Event coverage

For our synthetic construction worker Ahmed:

Daily income:             PKR 1,800
Working days/month:       24
Current savings:          PKR 18,000
Monthly expenses:         PKR 32,000
Health savings:           PKR 3,000
Dependents:               4
Estimated medical cost:   PKR 8,000
Days unable to work:      14

The resulting income interruption alone is:

$$ 1,800 \times 14 = 25,200\text{ PKR} $$

His estimated savings runway is approximately:

$$ \frac{18,000}{32,000/30} \approx 17\text{ days} $$

This demonstrates why even a relatively short health-related work interruption can create significant household pressure.

Financial Vulnerability Index

We developed a transparent Financial Vulnerability Index ranging from 0 to 100.

The index combines several financial resilience dimensions:

Component	Weight
Savings Runway	25%
Expense-to-Income Ratio	20%
Emergency Fund Gap	20%
Event Coverage	15%
Dependents	10%
Income Interruption Exposure	10%

One important design decision was to exclude medical risk probability from the Financial Vulnerability Index.

Health information should not silently become a credit score.

The Financial Vulnerability Index is therefore strictly informational.

It is not intended for:

Lending decisions
Credit scoring
Insurance pricing
Employment decisions
Worker eligibility decisions

Occupational Hazard Assessment

Different occupations involve different working environments.

SehatRozgar therefore includes a deterministic occupational hazard engine.

The prototype supports occupations including:

Construction Worker
Factory Worker
Warehouse Worker
Delivery Rider
Driver
Welder
Electrician
Agricultural Worker
Sanitation Worker
Security Worker

For example, a construction worker may encounter occupational hazards such as:

Falls
Dust exposure
Heat exposure
Noise exposure
Musculoskeletal strain

A delivery rider may encounter:

Traffic accidents
Heat exposure
Air pollution
Fatigue
Prolonged riding

A welder may encounter:

Fumes
Burns
Eye injury
Heat exposure
Respiratory exposure

These mappings provide occupational context only.

They are not predictions that an injury or illness will occur.

Worker Resilience Agent

Generative AI is used as an explanation layer, not as the source of truth.

The Worker Resilience Agent, powered by Featherless AI, receives structured outputs that have already been generated by the machine-learning model and deterministic engines.

For example:

{
  "health": {
    "risk_probability": 0.11,
    "risk_level": "low"
  },
  "financial": {
    "income_loss_14_days": 25200,
    "savings_runway_days": 17,
    "vulnerability_score": 63
  },
  "occupational": {
    "occupation": "Construction Worker",
    "hazards": [
      "falls",
      "dust exposure",
      "heat exposure"
    ]
  }
}

Featherless AI converts this structured information into understandable guidance.

The AI Resilience Advisor can provide explanations in:

English
Simple English
Roman Urdu

This allows technical information to be communicated in a more accessible way.

AI Safety Architecture

We deliberately restricted what the generative AI layer is allowed to do.

The AI Can

Explain model outputs
Summarize financial vulnerability
Explain occupational hazards
Prioritize existing recommendations
Suggest questions to discuss with a healthcare professional
Translate technical information into simpler language

The AI Cannot

Diagnose disease
Prescribe medication
Invent medical measurements
Invent probabilities
Modify ML predictions
Calculate financial metrics
Modify vulnerability scores
Determine creditworthiness
Price insurance
Make employment decisions

Before structured information is sent to the LLM, sensitive identifiers can be removed.

The architecture follows:

Structured Results
        |
        v
Data Sanitization
        |
        v
Featherless AI
        |
        v
Structured Explanation
        |
        v
User Interface

If Featherless AI becomes unavailable, the application can use a deterministic fallback explanation, allowing the core assessment system to continue functioning.

Workforce Analytics

SehatRozgar also demonstrates how resilience information could be analyzed at an aggregate level.

For the hackathon prototype, the workforce dashboard contains six completely synthetic worker profiles.

The dashboard visualizes information such as:

Number of synthetic workers
Workers with less than 14 days of savings runway
Average emergency preparedness
Financial vulnerability distribution
Occupation distribution
Occupational groups

Individual medical assessments are deliberately excluded from workforce-level analytics.

The purpose is to demonstrate aggregate resilience analysis without turning individual medical information into an employment-management tool.

Challenges We Ran Into

1. Connecting Healthcare and Finance Responsibly

The biggest conceptual challenge was combining two sensitive domains without producing a misleading universal "worker score."

A worker could have relatively low statistical health risk while simultaneously having extremely high financial vulnerability.

Combining these into one number would hide that distinction.

We therefore keep:

Health risk
Financial vulnerability
Occupational hazard
Emergency preparedness

as separate dimensions.

2. Deciding Where Generative AI Belongs

It would have been much easier to send all worker information to an LLM and ask it to generate an assessment.

We deliberately rejected that architecture.

Instead:

Machine Learning  → Statistical Prediction
Mathematics       → Financial Calculations
Rules             → Occupational Context
Generative AI     → Explanation

This makes the system more transparent, reproducible, and auditable.

3. Working With a Small Historical Dataset

The processed Cleveland dataset contains only 303 observations and was not collected specifically from daily-wage workers.

Instead of hiding this limitation, we expose:

Dataset size
Model metrics
Confusion matrix
Model comparison
Methodology
Limitations

directly inside the application.

The health component should therefore be understood as an empirical research demonstration, not a clinically validated system.

4. Separating Explainability From Causality

Another challenge was presenting model explanations responsibly.

A model may assign a strong contribution to a particular feature, but this does not mean that feature medically caused the outcome.

We therefore describe our visualization as model explainability, not causal medical explanation.

5. Making Technical Information Understandable

Metrics such as ROC-AUC, feature contribution, event coverage, savings runway, and emergency-fund gaps are useful technically but may not be intuitive for ordinary workers.

This motivated the Worker Resilience Agent.

Instead of replacing the analytical system with AI, we use AI to make the analytical system easier to understand.

What We Learned

The biggest lesson from building SehatRozgar was that adding AI to healthcare does not necessarily mean allowing AI to make more decisions.

Sometimes a stronger architecture gives AI less authority and a more focused responsibility.

We learned to separate responsibilities clearly:

Machine learning handles statistical prediction.

Mathematics handles financial calculations.

Rules provide occupational context.

Generative AI handles communication.

We also learned that financial resilience can dramatically change the meaning of a health event.

For someone with months of savings, a two-week interruption may be manageable.

For someone with only a few days of financial runway, the same interruption may create immediate household pressure.

That relationship became the central idea behind SehatRozgar.

Accomplishments We're Proud Of

Rather than building another medical chatbot, we created an end-to-end worker resilience intelligence system.

The prototype includes:

Real public healthcare data
Reproducible ML preprocessing
Multiple-model comparison
Holdout model evaluation
Model explainability
Deterministic financial calculations
Financial Vulnerability Index
Savings runway analysis
Income Interruption Simulator
Occupational hazard assessment
Emergency preparedness analysis
Synthetic workforce analytics
Featherless AI integration
AI-generated resilience explanations
Simple English explanations
Roman Urdu accessibility
Deterministic AI fallback
Interactive dashboards
Dedicated model-performance reporting
Transparent methodology and limitations

Most importantly, every component has a clearly defined responsibility.

Technology Stack

Frontend

React
TypeScript
Vite
Tailwind CSS
Recharts

Backend

Python
FastAPI
scikit-learn

Machine Learning

Logistic Regression
Random Forest
Gradient Boosting
UCI Heart Disease Dataset
Model explainability

Generative AI

Featherless AI
Structured prompting
Structured JSON outputs
Deterministic fallback guidance

Analytics

Financial resilience engine
Income interruption simulation
Occupational hazard rules
Workforce resilience analytics

Ethics and Privacy

SehatRozgar is currently a research prototype.

All worker profiles shown in the demonstration are synthetic.

The prototype:

Does not operate on real patients
Does not execute real financial transactions
Does not provide medical diagnoses
Does not prescribe treatment
Does not determine employment eligibility
Does not provide credit scores
Does not make lending decisions
Does not price insurance

The health-risk model is based on a small historical dataset and should not be interpreted as clinically validated for occupational worker populations.

Any future real-world healthcare deployment would require significantly stronger evidence, external validation, clinical oversight, privacy safeguards, bias evaluation, and appropriate regulatory review.

What's Next for SehatRozgar

SehatRozgar is currently a research prototype, but the architecture could be expanded significantly.

Future work could include:

Larger and more representative occupational-health datasets
Longitudinal worker health data
Income-volatility modeling
Country-specific social-protection information
Worker benefits and assistance discovery
Additional local languages
Voice-based accessibility
Mobile-first assessments
Wearable integrations
Occupational telemetry
Consent-based healthcare integrations
Longitudinal resilience monitoring
More sophisticated financial shock simulations

A future research direction could also study how different financial buffers change the economic impact of temporary health-related work interruptions across different occupations.

Our Vision

SehatRozgar started with one question:

What happens when a health event also stops someone's income?

The answer cannot be captured by a medical probability alone.

It involves health, occupation, income, savings, household responsibilities, and the worker's ability to absorb an unexpected interruption.

SehatRozgar demonstrates how these dimensions can be analyzed together while keeping their responsibilities separate and transparent.

By combining empirical machine learning, transparent financial modeling, occupational context, and accessible generative AI, SehatRozgar makes an often invisible consequence of illness easier to understand.

Health risk doesn't end at the hospital.

When health stops work, resilience becomes measurable.

Built With

ai
analytics
explainable
fastapi
featherless
financial
fintech
generative
healthcare
healthtech
llm
predictive
python
react
regression
scikit-learn
tailwind
typescript