How to run a Typing Tool analysis

The user types something like:

• "Build a typing tool for our customer segments."
• "Which questions do we need to classify new respondents into our segments?"
• "Create a short screener for our segmentation."

The platform recognises the intent and selects the Typing Tool.

The chat asks the user to identify:

• Segment label column — the column that contains the existing segment assignments (e.g. segment, cluster_label, customer_type)
• Candidate predictor columns — the survey questions that could be used to classify new respondents. Typically the user selects all non-demographic survey items.

The platform suggests mappings based on column names. The user confirms or adjusts.

Before the ladder runs, the chat asks how the client intends to use the deliverable:

• Excel workbook — for contact-centre or research teams who will score new respondents in Excel. Tests logistic regression and LDA only, so the workbook can compute predictions natively in formulas.
• Model bundle — for engineering teams embedding the classifier in their own systems. Tests up to fifteen classifier families and ships a pickled model with documentation.

This choice determines the classifier catalog used in evaluation.

• "Only respondents with complete segment labels"
• "Exclude segment F — it's too small to classify reliably"

The platform validates that enough labelled respondents remain per segment.

The chat shows a confirmation screen — segment label column, number of candidate predictors, output format, active filters, and the expected sample size and number of segments. The user confirms, and the evaluation phase runs.

The chat returns a ranked table of candidate question sets. For the Excel-workbook path this is every rung from 4 to 12 features × two classifiers (logistic and LDA) — 18 rows. Example:

The platform highlights its auto-suggestion — the smallest feature set that clears all three quality gates (accuracy ≥ 75%, macro-F1 ≥ 70%, ECE ≤ 1.00), with logistic preferred over LDA on ties — and explains the recommendation. For the Model-bundle path, a fixed feature set is screened and the ladder shows up to fifteen classifier rows instead.

The user picks the rung they prefer. Common considerations:

• "Six questions is short enough for our contact centre script — go with FS-6."
• "We can afford ten questions in the online form — let's get the extra accuracy."
• "Four questions is too inaccurate — bump to six."

The platform fits the final model on the full dataset, generates scenario predictions (what would the model predict for typical Low / Mid / High respondent profiles?), and produces — depending on the output format chosen at step 3:

Excel-workbook path:

• An Excel typing-tool workbook — the operational deliverable. Drop in new respondent answers, get back predicted segment and confidence, all computed natively in Excel formulas.
• A feature-set ladder CSV — the full evaluation table for reference.
• A written report with performance summary and scenario predictions.

Model-bundle path:

• A trained model bundle (joblib pickle) — production-grade classifier for embedding in CRM, scoring APIs, or batch pipelines.
• A generated README.md documenting the chosen classifier, its features, and the evaluation metrics.
• A generated load_classifier.py loader script showing how to score new respondents in Python.
• A feature-set ladder CSV and a written report.

• "What if we drop question Q14 — it's hard to ask in a call centre?"
• "Can we try with just 4 questions instead of 6?"
• "Show me how accurate it is for Segment C specifically."

The user can re-run the evaluation with adjusted parameters without re-uploading or starting from scratch.