Methodology
From source reports to readable text, themes and evidence you can explore.
Python + WordPress API: collect reports
Python and the WordPress API downloaded public reports from Isa’s office website and saved links to each source.
Beautiful Soup: clean the text
Beautiful Soup removed webpage code so we could read and analyse the report text.
Source checks: avoid repeats and wrong totals
We removed exact copies and checked figures against their source reports, keeping plans separate from reported delivery and avoiding overlapping beneficiary totals.
TF-IDF + NMF: find themes
scikit-learn uses TF-IDF to highlight distinctive words and NMF to find 8 recurring themes in the selected English reports.
Manual categories: label activities
We assigned activity categories by reading the reports; one activity can belong to several categories.
Phrase counts: build the word cloud
A script counts each phrase once per report for the concept cloud; the category view counts activities. Larger labels mean more records.
Sources and guided chat: explain the findings
The dashboard links figures and connections to reports; the chatbot follows prepared questions. The reports do not independently prove impact or cover his whole career.
How AI agents helped our workflow
The creator worked with AI coding agents to turn the idea into a working website. They helped write collection and analysis scripts, build the pages and guided questions, prepare draft translations, and refine the design through feedback. Focused agent tasks also helped examine research options and check parts of the work.
The owner set the direction and corrected the wording and presentation. Agents ran automated content, calculation and browser checks, with source reports kept alongside the claims. This support does not replace the original evidence or certify translations. The visitor-facing chatbot follows prepared questions and answers; it does not call a live AI agent.
Technical settings
Settings for the selected English-text model. Technical parameter names are retained in English.
- Features
- TF-IDF; unigrams + bigrams; min_df=3; max_df=0.85; sublinear_tf=True
- Model
- NMF; 8 components; nndsvda; coordinate descent; random_state=42
- Convergence
- max_iter=600; tol=0.0001; selected fit: 839 features, 117 iterations
Sources and implementation
Inspect the implementation and saved methodology records: