Intelligent Search · Advertising · production
MediaX
Semantic and geospatial OOH recommendation engine.
LLM-driven system that transforms natural-language campaign requirements into structured geographic intent, retrieves inventory, and ranks placements across 290+ UAE categories in under five seconds.
Geospatial OOH media discovery
INTENT
My role
Applied AI / Search Systems
- AI Architecture
- Retrieval
- Backend Engineering
- Data Pipelines
- Productionization
Period / 2025
Overview
LLM-driven system that transforms natural-language campaign requirements into structured geographic intent, retrieves inventory, and ranks placements across 290+ UAE categories in under five seconds.
Problem
OOH campaign planning requires translating vague location language into validated geo-intent and ranking inventory against footfall, traffic, POI density, and proximity—manually and inconsistently.
System
An intent parser plus geospatial retrieval and weighted ranking pipeline with enrichment, caching, retries, and idempotent upserts for reliable recommendation responses.
System anatomy
Architecture
Intent → geo resolution → retrieval → weighted ranking.
- Intent parser
- Geo/POI resolution
- Ranking engine
- API enrichment
Architecture
Natural-language campaigns flow through intent parsing, geo/POI resolution, semantic validation, geospatial retrieval, and a weighted ranking engine into inventory recommendations.
- NL Campaign → Intent Parser
- Intent Parser → Geo + POI Resolution
- Geo + POI Resolution → Semantic Validation
- Semantic Validation → Geospatial Retrieval
- Geospatial Retrieval → Ranking Engine
- Ranking Engine → Inventory Recommendation
Engineering decisions
Structured geo-intent extraction
Campaign language is ambiguous. Parsing into validated locations and POI types makes retrieval deterministic and auditable.
Weighted business ranking
Inventory quality is multi-signal. Explicit weights (30/30/20/20) encode business priorities instead of opaque model scores alone.
Caching, retry/backoff, idempotent upserts
External APIs and enrichment steps fail. Retry/backoff, caching, conflict resolution, and idempotent writes keep the pipeline production-safe under load.
Reliability / production
- Retry / backoff
- Caching
- Conflict resolution
- Idempotent upserts
- Location semantic validation
Stack
Outcomes
- Natural-language campaigns mapped to structured geo-intent and POI types.
- Weighted ranking across footfall, traffic, POI density, and proximity.
- Optimized recommendations returned in under five seconds.
Links
Product visuals
