Escucha Comunitaria
The problem
Cartagena has sectoral roundtables, perception surveys, field-recorded testimonials. Dozens of documents that describe precisely what the community needs. None of it is systematically cross-referenced with the intervention proposals that land on a policymaker’s desk. The voices exist; the problem is that nobody consults them.
At the CTW Hackathon in Cartagena (August 13, 2026), team D&A chose that problem. Six hours, two people, 21 teams competing. We placed second.
The corpus
Data was collected between the hackathon day and the days leading up to it, combining open sources and direct field collection:
- 1,148 fragments from the 2026 Citizen Perception Survey of Cartagena (1,150 sources)
- 271 fragments from sectoral roundtable minutes from Puerto Colombia (26 sources)
- 38 fragments from direct testimonials recorded in the field and transcribed with Whisper (7 sources)
In total: 1,457 citizen voice fragments, 1,183 sources, 79 emerging topics, covering the 2023–2026 period.
No testimony is published without explicit consent. Field recordings are ingested marked private by default; only a human confirmation makes them visible.
How it works
Information goes through four stages before it becomes queryable:
- Transcription (Whisper) — field audio is converted to text automatically.
- Segmentation (Gemma 4 12B + Claude Haiku) — each document is split into minimal thematic units; local model first, API as fallback.
- Semantic indexing (Voyage voyage-4) — each fragment becomes a 1,024-dimension vector. Search is by meaning, not keyword.
- Topic discovery (BERTopic + UMAP + HDBSCAN) — fragments are grouped by semantic proximity. No topics were defined in advance.
With the corpus indexed, there are three ways to query it:
Contrast a proposal. Paste any social program proposal. The system breaks it into four dimensions (population, intervention, duration, expected outcome) and contrasts each one against the corpus with literal citations. The result can be “supported”, “tensioned”, or “no data”.
Explore topics. The topics that emerge from the corpus without anyone defining them beforehand, ordered by volume of evidence and with their distribution across source types.
Conversational query. Ask in natural language. The system retrieves the most relevant fragments and responds citing each source. Filterable by municipality.
The team
D&A: Dorelys Martínez Suárez (AI, RAG, pipeline) and Alex Martínez (data collection, field interviews). The division was intentional: while I built the system, Alex did the harder job — getting Cartagena residents to speak honestly.
Current status
Under continuous development. The system is now being applied to Puerto Colombia, Atlántico, our home municipality, expanding the corpus with new local sources.
The full story of the process, including the fieldwork, is in this blog post.