Email NLP Classifier & NER -- 2
Budget: ₹600 – ₹1,500 INR
I have a large, continually growing collection of emails that needs to be processed automatically. The goal is twofold:
1. Classify each email into predefined business categories with high accuracy.
2. Extract relevant entities (names, dates, IDs, product references, etc.) from the same messages.
You will own the entire machine-learning workflow. That means cleaning and exploring the raw email text, crafting useful features, training and tuning your models, and packaging the final solution behind an API that I can call from our existing back-end. Python is a must, and I’m comfortable with either TensorFlow or PyTorch for the deep-learning components—use whichever lets you move fastest. Traditional techniques with Scikit-learn are welcome wherever they make sense.
Because the email stream is live, the models have to run efficiently at scale and be easy to retrain when new data arrives. A clean, well-documented repo, reproducible training scripts, and a straightforward deployment path to a common cloud platform will be part of the hand-off.
Deliverables (all discrete and specific):
• Pre-processing pipeline that ingests raw emails and outputs tokenized, ready-to-train data
• Trained text-classification model and NER model with evaluation reports (precision, recall, F1)
• Inference service exposed as REST or gRPC endpoint, containerised (Docker) for cloud deployment
• Read-me style documentation covering setup, retraining, and expected input/output formats
Acceptance criteria: models must achieve an F1 score ≥ 0.9 on a held-out test set we’ll agree upon, spin up in under 500 ms per request, and handle at least 50 requests per second without degradation.
If this first phase goes smoothly, I have plenty of follow-on work—additional languages, sentiment analysis, even computer-vision tie-ins—so think of this as the start of a longer collaboration.
1. Classify each email into predefined business categories with high accuracy.
2. Extract relevant entities (names, dates, IDs, product references, etc.) from the same messages.
You will own the entire machine-learning workflow. That means cleaning and exploring the raw email text, crafting useful features, training and tuning your models, and packaging the final solution behind an API that I can call from our existing back-end. Python is a must, and I’m comfortable with either TensorFlow or PyTorch for the deep-learning components—use whichever lets you move fastest. Traditional techniques with Scikit-learn are welcome wherever they make sense.
Because the email stream is live, the models have to run efficiently at scale and be easy to retrain when new data arrives. A clean, well-documented repo, reproducible training scripts, and a straightforward deployment path to a common cloud platform will be part of the hand-off.
Deliverables (all discrete and specific):
• Pre-processing pipeline that ingests raw emails and outputs tokenized, ready-to-train data
• Trained text-classification model and NER model with evaluation reports (precision, recall, F1)
• Inference service exposed as REST or gRPC endpoint, containerised (Docker) for cloud deployment
• Read-me style documentation covering setup, retraining, and expected input/output formats
Acceptance criteria: models must achieve an F1 score ≥ 0.9 on a held-out test set we’ll agree upon, spin up in under 500 ms per request, and handle at least 50 requests per second without degradation.
If this first phase goes smoothly, I have plenty of follow-on work—additional languages, sentiment analysis, even computer-vision tie-ins—so think of this as the start of a longer collaboration.