Implemented
Lowercase word tokenization, stop-word removal, multinomial Naive Bayes, Laplace smoothing, log-space scoring, known-vocabulary gating, held-out evaluation, confusion matrix and JSON export. Train/test ID overlap is blocked.
Independent engineering sampleMethod & limitsMHTECHIN / JUNIOR AI ENGINEER
A small, executable machine-learning workflow for enterprise support intake. Train a text classifier, test what it has not seen, and keep uncertain tickets with a person.
02 / Held-out evidence
Review is an explicit expected label for four ambiguous, unrelated or empty cases. F1 includes review. This tiny, hand-written set is deliberately simple and cannot establish deployment quality.
The held-out set is never used to fit the model.
03 / Try an unseen ticket
Type a support request. The model reveals its scores, vocabulary coverage and token evidence relative to the runner-up class.
Train a model first. Your text stays in this page and is not saved or transmitted.
04 / Implementation & limits
Lowercase word tokenization, stop-word removal, multinomial Naive Bayes, Laplace smoothing, log-space scoring, known-vocabulary gating, held-out evaluation, confusion matrix and JSON export. Train/test ID overlap is blocked.
27 configurations · 1,466 assertions. Synthetic tests check score normalization, routing gates, empty/unseen input, duplicate IDs and train/test overlap. Default toy-set results are not an estimate of customer accuracy.
FastAPI service, authenticated intake adapters, private labeled ticket data, time/customer-separated evaluation, calibration, multilingual testing, drift monitoring and reviewer feedback. No external APIs, database or automatic ticket writes are connected.