Hasan Ali

We were provided with coordinates of farmer lands and drone survey data (an ECW file) for an area.

TASK
• Extract land images for every farmer.
• Identify crops on those images.
• Estimate the output of each crop for each farmer.

SOLUTION
QGIS was used to process the ECW file. I built a Python script to process the data and extract farmer land, using the coordinates from that ECW file and saving it in PNG format.

After that, an image classifier with a sliding window was used to detect the crops in the image. Later in the project, we used image segmentation to identify the total area occupied by the crop. Deep learning and image processing were used to estimate the crop yield for a particular farmer from that area.

Objective: Build a product that takes any given document (PDF or image) and blurs the personally identifiable information (PII) on it.

The application could be used to:
1. Blur critical information on government-issued documents.
2. Blur the patient's name and other personal details on x-rays and other medical reports to make the diagnosis independent of the patient.

Solution: The product used custom object detection to detect text, faces, and document boundaries. Using boundaries, the main area was extracted and sent to OCR. The next step was to blur the areas in the image that show any critical PII.

Task: Detect all anomalies and metal loss inside oil pipelines.

Solution: Oil pipelines have multiple anomalies, such as weld, sleeve, and T, with very high imbalance. First, we created and engineered a training dataset with a proportionate distribution of anomalies and normalities. Then, we trained a fully connected CNN and ML algorithm solution along with some post-processing. All these measures were used to ensure that we didn't miss any anomalies. So, overall, it was kind of a double-surety architecture.

For metal loss detection (highly imbalanced data): Again, we created and engineered a training dataset and used an image segmentation (modified U-net) with a custom loss function.

Impact: We automated the anomaly detection process and saved 1,000+ person-hours and $100,000+ every week.

Inventory forecasting for a retail client with 500+ SKUs, including an ensemble of many time series-based models (like ARIMA), probability-based models, and a transformer for time-series forecasting with explainability. While creating the model, I added explainability, so the model predictions came with a confidence interval and an explanation of the reason for the predictions.

The LLM-based Support Assistant is a cutting-edge solution that provides students with quick and accurate answers to their queries using Large Language Models (LLMs) such as LlamaIndex and OpenAI GPT-3.5-turbo. This assistant leverages the power of natural language processing to comprehend and respond to a wide range of questions related to academic subjects, assignments, exams, and more.