
## Overview

The Travel Price Predictor is a sophisticated time-series forecasting system that analyzes historical pricing data, seasonal patterns, and market signals to predict optimal booking windows for flights and hotels.

## Business Problem

Thrillers Travels needed to:

- Maximize revenue by suggesting optimal pricing
- Help customers find the best booking times
- Reduce inventory waste from poor timing
- Compete with larger OTAs using data-driven insights

## Data Pipeline

We built a robust data pipeline processing over 10 million data points daily:

### Data Sources

- Historical booking data (5 years)
- Competitor pricing (scraped hourly)
- Flight capacity data
- Seasonal events calendar
- Economic indicators

### Feature Engineering

```python
def create_features(df):
    df['day_of_week'] = df['date'].dt.dayofweek
    df['days_until_departure'] = (df['departure'] - df['date']).dt.days
    df['is_holiday_season'] = df['date'].apply(check_holiday)
    df['competitor_price_ratio'] = df['our_price'] / df['competitor_avg']
    return df
```

## Model Architecture

We employed an ensemble approach combining:

1. **Prophet**: For capturing seasonality and trends
2. **XGBoost**: For feature-based predictions
3. **LSTM**: For sequential pattern recognition

The final prediction is a weighted average based on recent performance.

## MLOps Pipeline

The entire system runs on AWS SageMaker with:

- **Training**: Weekly retraining on fresh data
- **Inference**: Real-time predictions via API
- **Monitoring**: CloudWatch dashboards for model drift
- **A/B Testing**: Continuous evaluation against baseline

## Results

- **15% increase** in revenue from dynamic pricing
- **23% improvement** in booking conversion rates
- **$50K+ saved** annually from optimized inventory
- **Real-time predictions** in under 100ms

## Key Insights

1. Booking patterns vary significantly by destination type
2. Competitor pricing is highly predictive within 48 hours
3. Social media signals provide early demand indicators
4. Weather forecasts impact last-minute bookings



## Overview

Activity Vision is a privacy-first productivity monitoring solution that uses computer vision to understand work patterns without invasive keystroke logging or continuous screen recording.

## The Problem

Remote work monitoring tools face a dilemma:

- Too invasive = employee distrust
- Too lax = no actionable insights

We needed to find a balance that respects privacy while providing useful productivity data.

## Technical Solution

### Snapshot Analysis

Instead of recording screens, we take periodic snapshots and immediately process them locally:

```python
class ActivityClassifier:
    def __init__(self):
        self.ocr_engine = TesseractEngine()
        self.classifier = load_model('activity_classifier.h5')

    def analyze_screenshot(self, image):
        # Extract text for app identification
        text = self.ocr_engine.extract(image)

        # Classify activity type
        features = self.extract_features(image, text)
        activity = self.classifier.predict(features)

        # Delete image immediately
        del image

        return activity, confidence
```

### Privacy by Design

- Images are processed locally and immediately deleted
- Only activity categories are sent to the server
- No keystroke logging or clipboard monitoring
- User can pause tracking anytime

### Activity Categories

The model classifies work into categories:

1. **Coding**: IDE usage, terminal commands
2. **Communication**: Email, Slack, meetings
3. **Research**: Documentation, web browsing
4. **Design**: Creative tools, presentations
5. **Administrative**: Spreadsheets, documents

## Model Training

We trained on 50,000+ anonymized screenshots with:

- Multi-task learning (app + activity type)
- Data augmentation for screen variations
- Transfer learning from ImageNet

## Edge Deployment

For privacy and latency, the model runs entirely on-device:

- Model size: 15MB (quantized)
- Inference: <100ms on CPU
- Memory: <200MB RAM

## Results

- **5,000+ active users** across 50 companies
- **100ms average** processing time
- **94% accuracy** in activity classification
- **Zero privacy complaints** in 2 years

## Future Improvements

- Attention pattern analysis (with consent)
- Meeting engagement scoring
- Burnout prediction alerts
- Integration with project management tools


All Projects

Travel Price Predictor

Activity Vision