Enterprise Email Classifier API
Budget: ₹2,000 – ₹3,000 INR
I have an anonymised set of internal company emails that must be routed automatically to the correct department—Human Resources, Finance or IT Support. I am looking for an engineer who can take the project from raw data through to a Docker-ised REST service.
The workflow I expect is:
• Data preparation: strip any residual PII, apply consistent labelling for the three target departments, and document the pipeline so it can be rerun when fresh mail arrives.
• Modelling: start with a TF-IDF + SVM baseline, then fine-tune transformer models (BERT or RoBERTa via Hugging Face). Compare approaches and capture precision, recall and F1 for each class. An error analysis explaining common misclassifications is essential.
• Deployment: package the best model behind a FastAPI endpoint, containerised with Docker so I can spin it up on any Linux box.
• Documentation: a clear setup guide, plus a short “retraining cheatsheet” outlining how to refresh labels, retrain and redeploy.
Deliverables
– Cleaned & labelled dataset (no PII)
– Jupyter notebooks / scripts for both baselines and transformer models
– Evaluation report with metrics and error analysis
– Production-ready FastAPI service in a Dockerfile
– README covering environment setup, inference usage and retraining steps
– All source code in a private Git repo or zipped archive
Acceptance criteria
1. F1-score ≥ 0.90 averaged across HR, Finance and IT Support on a held-out test split.
2. Docker image builds and serves predictions on “/predict” within 500 ms per email on CPU.
3. Documentation enables a fresh setup (from clone to inference) in under 30 minutes.
Please share links or brief summaries of previous NLP or text-classification projects—especially any that involved FastAPI or Hugging Face. I’m ready to begin immediately and aim to wrap up within four weeks.
The workflow I expect is:
• Data preparation: strip any residual PII, apply consistent labelling for the three target departments, and document the pipeline so it can be rerun when fresh mail arrives.
• Modelling: start with a TF-IDF + SVM baseline, then fine-tune transformer models (BERT or RoBERTa via Hugging Face). Compare approaches and capture precision, recall and F1 for each class. An error analysis explaining common misclassifications is essential.
• Deployment: package the best model behind a FastAPI endpoint, containerised with Docker so I can spin it up on any Linux box.
• Documentation: a clear setup guide, plus a short “retraining cheatsheet” outlining how to refresh labels, retrain and redeploy.
Deliverables
– Cleaned & labelled dataset (no PII)
– Jupyter notebooks / scripts for both baselines and transformer models
– Evaluation report with metrics and error analysis
– Production-ready FastAPI service in a Dockerfile
– README covering environment setup, inference usage and retraining steps
– All source code in a private Git repo or zipped archive
Acceptance criteria
1. F1-score ≥ 0.90 averaged across HR, Finance and IT Support on a held-out test split.
2. Docker image builds and serves predictions on “/predict” within 500 ms per email on CPU.
3. Documentation enables a fresh setup (from clone to inference) in under 30 minutes.
Please share links or brief summaries of previous NLP or text-classification projects—especially any that involved FastAPI or Hugging Face. I’m ready to begin immediately and aim to wrap up within four weeks.