Structured extraction before product logic
The model produces fields such as location, severity, summary, actors, and affected people. That gives downstream relevance and notification steps a stable shape instead of making them parse prose.
← Back to selected work
Truquito
A news and events product that helps people in Ecuador find relevant information about places near them.
RoleFounder & sole technical owner
FocusReact Native · Supabase · n8n
01Context
The problem
National news is abundant, but local relevance is hard to discover. Truquito turns a stream of public sources into structured, location-aware signals that can reach people nearby.
My contribution
I designed, built, deployed, and operate the product end to end. That includes the mobile experience, data model, automated workflows, model integration, geolocation, notifications, and the day-to-day work of keeping sources healthy.
02System view
How the pieces connect
A conceptual view of the system, recreated from the public-safe project description.
01RSSNational sources
02ExtractStructured fields with an LLM
03LocatePlaces, coordinates, relevance
04StoreNews and event records
05NotifyNearby users
03Outcome
The result is a production product with an automated processing loop and a deliberately lean operating model. It can ingest, interpret, locate, and route relevant items without regular manual editorial work.
15national RSS sources in the ingestion design
10 minsource polling cadence
~200articles processed per day in the current notes
04Engineering notes
Managed services for a small operating footprint
Supabase, Vercel, n8n, and external location services keep the product practical to run as a solo technical owner. The architecture favors leverage over infrastructure that would require constant attention.
Embeddings as preparation, not a public claim
News and events receive OpenAI text-embedding-3-small vectors to prepare future intent-based discovery. Semantic search is not presented as a public feature already launched.
05Scope & limits
Stack
What I would keep visible
- The most common operational failure is a source discontinuing its RSS feed; it is repaired or temporarily disabled through n8n.
- Consent, retention, encryption, and access-control details for personalized location data still need to be documented before publication.
- Adoption and operating-cost figures are intentionally not presented here as current public metrics until they are confirmed.
The most important engineering decision is not treating the model as the product. The useful product is the complete loop around it: inputs, structured output, location, relevance, delivery, and recovery when a source changes.
Explore another case