Webinar materials & announcements
AI kategoryzacja wydatków, Sesja przez Sesja
Ustrukturyzowany learning dla professionals building real classification systems
Each webinar at Setawoy jest stworzony around a specific technical problem - nie a broad survey. Sesje cover model selection, dane labelling approaches, edge-case handling, i production integracja, w enough depth do być immediately applicable.
Sesja catalogue
Co the sesje cover
Practical Webinary na the mechanics zautomatyzowany expense classification - od raw transaction dane do a working categorisation pipeline.
Most classification problems w finanse fail at the dane layer, nie the model layer. Ten sesja examines jak transaction descriptions vary w całym banks, merchants, i currencies - i dlaczego a single regex-based approach breaks down beyond a few hundred categories.
- Dlaczego merchant Imię i nazwisko normalisation matters przed dowolne model training
- Comparing rule-based, ML, i hybrid approaches w całym dataset sizes
- Handling ambiguous transactions że fit multiple categories
- Evaluation metrics że actually reflect categorisation quality w practice
Annotation quality determines model ceiling. Ten sesja walks poprzez practical labelling strategies dla expense dane - w tym jak do handle disagreements między annotators kiedy a transaction could reasonably belong do two categories.
- Designing a category taxonomy że scales beyond 200 labels
- Active learning do reduce annotation effort na large datasets
- Detecting i correcting label noise po the fact
Large language modele są nie automatycznie the right narzędzie dla expense categorisation. Ten sesja compares fine-tuned BERT-style modele against gradient-boosted classifiers na real transaction dane, z latency i koszt jako explicit constraints.
- Kiedy a smaller model outperforms a larger one na ustrukturyzowany text
- Feature engineering dla traditional ML na finansowy descriptions
- Inference speed trade-offs at production transaction volumes
A model że works w a notebook does nie automatycznie work w production. Ten sesja covers the gap - batch vs. real-time inference, fallback logic kiedy confidence jest Niski, i jak do monitor category drift jako merchant behaviour changes over time.
- Structuring a confidence threshold dla human-in-the-loop przegląd
- Logging predictions w a way że supports future retraining
- Detecting distribution shift w incoming transaction dane
International expense dane introduces classification problems że domestic datasets rarely expose. Ten sesja examines jak do handle merchant names w non-Latin scripts, currency-dependent category conventions, i transactions że span multiple koszt centres.
- Transliteration i normalisation strategies dla global merchant dane
- Category schema alignment w całym different regional accounting standards
- Confidence calibration kiedy training dane jest sparse dla a currency region
Co participants take away
Sesje są recorded i materials stay accessible po each live date. The goal jest nie do watch a presentation - it jest do leave z a concrete next step że fits Twoje current project.
Mentorship sesje following each webinar allow participants do apply concepts do their specific dataset i get direct feedback na wdrożenie decyzje.
The sesja na model selection changed jak I framed the problem entirely. I had been trying do fine-tune a large model na 4,000 labelled transactions i getting mediocre wyniki. Switching do a gradient-boosted approach z better feature engineering cut my error rate noticeably i the inference time dropped od seconds do milliseconds.