Interning at Prediction3d

My time as an AI and Machine Learning intern at Prediction3d was a massive period of growth. Transitioning from academic coursework to production models is a steep learning curve, but incredibly rewarding.

Why real-world ML feels different

In coursework, the model is the center of attention. In production, reliability, observability, and data quality dominate engineering time.

That shift in priorities was the biggest mindset change from internship work.

The Gap Between Theory and Practice

In school, you are typically handed a clean dataset (like MNIST or Titanic) and told to build a model. In the real world, the data is messy, incomplete, and sometimes explicitly wrong.

Data Cleaning is 80% of the Job

I spent far more time writing Pandas and Numpy scripts to clean and normalize data pipelines than I did actually tuning hyperparameters in PyTorch or TensorFlow.

Model Optimization techniques

We had models that performed well contextually but were too slow or resource-heavy for production edge deployment. I learned several optimization techniques:

1. Quantization: Converting FP32 weights to INT8, significantly reducing model size with minimal accuracy loss.
2. Pruning: Removing near-zero weights from the network to compress the architecture.
3. ONNX Export: Getting models out of native PyTorch and into ONNX runtime for cross-platform inference speedups.

Productionizing ML is less about finding the exact perfect theoretical architecture, and more about finding the "good enough" architecture that fits into memory and runs in <50ms.

Takeaways

The most valuable thing I learned wasn't a specific framework API, but rather the intuition of diagnosing a badly behaving model. Understanding why a gradient is vanishing or why the loss curve resembles a chaotic heartbeat is what separates a practitioner from someone just copying tutorials.