AB-NODA

1. Problem

   News plus a language model detects what people are discussing. Strategy needs to know what is changing: patents filed, papers published, capacity committed, rules issued, production moving. Those live in different sources with different contracts.

2. Capability
   Specialized Data PipelinesKnowledge InfrastructureAgentic AI
3. Architecture

   One FastAPI service, one PostgreSQL schema and one port per domain, no cross-domain imports. PostGIS, pgvector and DuckDB where each fits. RabbitMQ for async work. MINERVA: 29 registered sensors (OpenAlex, EPO OPS, INPI, EIA, IRENA, ANP, JODI, SODIR, Baker Hughes, twelve trade-press feeds) produce two evidence classes: documents through a single LLM extraction contract, indicators differenced deterministically with no model. Signals are dated, attributed, typed observations backed by exactly one piece of evidence, with deterministic ids so re-collection deduplicates. TrendScore measures movement; Confidence measures how much the evidence entitles the claim; they are never merged. Divergence between sources is reported, never resolved.

4. Demonstrator

   Demonstrations on request.

5. Business impact
   • Entities resolved against official registries, with refusal as a recorded outcome
   • Signals with provenance from paragraph to source
   • A vertical configured in weeks, not years: oil & gas first

Modules