Get requests to
the right team.
Turn a free-text request into a reviewable next step. A practical case in understanding the work, implementing the model and planning the handover.
Independent implementation · Measured public-data evaluation
The next team needs a clear starting point.
A support request arrives in the customer’s own words. Someone has to identify what it concerns, decide where it belongs and avoid passing a confident mistake down the queue.
What I built
A trained text-classification pipeline with a review policy and a record of every test prediction. It makes the suggested category, uncertainty and mistakes inspectable before any routing decision.
“I still have not received my new card, I ordered over a week ago.”
Matches this dataset label. Source request #2, BANKING77 official test set. This is a saved model result.
- 01Read the request
- 02Suggest a category
- 03Keep review in the loop
- 04Inspect the mistakes
88.0% correctly classified across 3,080 held-out requests. The review-selected suggestions still include 111 mistakes; those errors remain available to inspect.
Independent implementation on attributed public data. The code, source records and complete evaluation are inspectable; confidential freelance work is described separately.
Inspect the evaluation belowWhat the build covers.
Data preparation, a trained model, a review policy and an interface for examining every result.
Prepare the data
Audit duplicates and labels, separate learning from validation, and keep the final test examples outside model selection.
Implement the decision
Train the 77-category classifier, compare settings and choose a review threshold using validation data.
Make the result inspectable
Retain every prediction, alternative category, review decision and error alongside the Python script and source records.
This implementation exposes saved held-out predictions. It does not connect to a client inbox or classify arbitrary new messages in the browser.
Python · scikit-learn · NLP · data quality · model evaluation · React · TypeScript
Project scope and implementation notes