Stakeholders
Media, institutions, employeesPublic conversation
News, blogs, social and TVInstitutional agenda
Official gazettes, chambers and regulatorsAssets and sponsorships
Presence at events and agreementsDigital footprint
Traffic, geolocation, browsingIntelligence. Not artificial.
Structure
Ordered data and comparable KPIsDiagnosis
Why reputation shiftsDecision
Information ready to act onTrust
Legitimate results, no hallucinationsSource verification
We audit the provenance of every informational input so the basis of the analysis is sound and truthful.
Human judgement
Unlike automated models, we apply cultural and political context to the data.
Decision guarantee
The final product is not just a report: it is a validated roadmap for critical decisions.
Monitoring is not a folder of clippings.
Newsrooms, institutions and comms teams watch dozens of sources by hand: feeds, official gazettes, television, social platforms. Every source arrives in its own format, every tool lives outside your control, and nothing connects to anything else.
Scattered sources
RSS, gazettes, TV, social and search, each in its own shape
No shared model
Nothing comparable across outlets or languages
Manual reading
Analysts spend the day filtering instead of deciding
KatanServant collapses that into a single pipeline. One canonical format. Your infrastructure.
Five stages, one canonical format.
Every document travels the same path, whatever its origin. Field mapping is declarative, and the system proposes rules that a person approves — it never guesses silently.
extract
Nine connectors: RSS, Google News, GDELT, MediaCloud, Google Search, X, YouTube, Acceso360 and legal gazettes
derive_text
PDF, audio and video into canonical text
normalize
Field mapping to NewsML-G2, enriched and de-duplicated
process
Entities, IPTC media topics, language
analyze
Threat classification, sentiment and cascading summarisation
The pipeline is a service, not a button.
The scheduler does not wait to be asked. It polls sources on its own cadence, runs the stages and writes to PostgreSQL with a MongoDB backup, so a failure in the backup never stops ingestion.
Threat taxonomy
14 categories × 5 levels
Story clustering
Jaccard with category weighting
Summarisation
Ollama → Groq → OpenRouter → extractive
Persistence
PostgreSQL primary, MongoDB backup
It runs where you decide.
Multi-tenant by design: tenants, projects, sources and documents, with five hierarchical roles resolved on every request. Federated sign-in against your own identity provider, with no automatic provisioning: an identity that is valid elsewhere does not grant access here.
Multi-tenant
Tenants → projects → sources → documents
Roles
Five levels, from viewer to super admin
Identity
OIDC against your provider, or local sign-in
Deployment
Docker on your servers or private cloud
What people ask before deciding.
Nine connectors today: RSS, Google News, GDELT, MediaCloud, Google Search, X, YouTube, Acceso360 and official legal gazettes. Adding one means writing a connector and a set of mapping rules, not changing the pipeline.
On your infrastructure. The stack is Docker Compose — API gateway, front end, Redis — against a PostgreSQL you control, on your own servers or a private cloud. There is no dependency on our hosting.
Two ways, side by side: local sign-in, or federated sign-in against your own identity provider over OIDC. Federated identities are not provisioned automatically: an account that is valid at your provider still needs to exist here before it grants access.
It travels five stages — extract, derive_text, normalize, process, analyze — and ends in a canonical NewsML-G2 shape. Field mapping is declarative: the system proposes rules and a person approves them, so it never guesses silently.
Summarisation runs a cascade: a local Ollama model first, then Groq, then OpenRouter, and an extractive fallback if all fail — so no single provider can stop the pipeline. Threat classification uses a 14-category by 5-level taxonomy over Spanish and English keywords.
Ingestion is language-agnostic. The processing stage currently has full support for Spanish, English and French; documents in other languages are stored and normalised, but entity and topic extraction is weaker.
Want to see it running on your own sources?
The fastest way to judge the platform is to point it at the sources you already watch and read what comes out.
Sign in