Big Data and the Television Sector

June 10, 2025 · Baoss
BAOSS IT consulting: agentic Data Mesh for the television sector

Big Data TV 2025: Generative AI Solves Fragmentation

In June 2015 we published on this blog how «TV Everywhere» and the proliferation of screens were breaking traditional measurement models. Ten years later, the problem is not a lack of data — it is an excess of noisy signals, the disappearance of third-party cookies, the entry into force of the European AI Act and the need to activate insights in milliseconds, not in monthly reports. At BAOSS IT consulting we solve it today with an agentic data architecture. The impact of AI-agent disruption explains the sector context.

The real 2025-2026 problem: fragmentation + privacy + latency

The Spanish audiovisual landscape in 2025 presents three vectors of tension that no conventional analytics suite solves:

  1. Extreme fragmentation: 8.2 connected devices per household (INE 2024), 14 active SVOD/AVOD/FAST platforms on average, and 68% of consumption on CTV that exposes no persistent IDs.
  2. Regulatory privacy: GDPR + ePrivacy + AI Act (full application August 2026) prohibit persistent profiles without explicit consent and demand algorithmic explainability.
  3. Decision latency: Programmatic CTV buying closes auctions in <50 ms; batch attribution models (T+1, T+7) arrive too late to optimise bid, frequency or dynamic creative.

The result: broadcasters and platforms lose between 22% and 35% of the value of their inventory due to the inability to match audience + context + intent in real time (source: IAB Spain / PwC Media Outlook 2025).

Why 2015-2022 architectures fail today

The «raw → curated → served» data lakes designed for descriptive BI do not support:

  • Multimodal data: video (frames, audio, subtitles), player telemetry, app events, social listening, enriched EPG.
  • Edge inference: running embedding models (e.g. jina-embeddings-v3 or nomic-embed-text-v1.5) on the CDN or set-top-box itself to pseudonymise before data leaves the perimeter.
  • Agent orchestration: flows where a planner agent (LangGraph / CrewAI / AutoGen) decides which tool to invoke: RAG over the catalogue, LLM-generated SQL, measurement API calls, bid simulation.

The BAOSS IT consulting 2025 architecture: an “Agentic Data Mesh” with privacy by design

Over the last 18 months we have deployed, at three large Spanish audiovisual groups, an architecture that replaces centralised ETL with federated data products governed by data contracts (OpenLineage + Great Expectations) and exposed via MCP (Model Context Protocol) to a layer of specialised agents.

Ingestion layer: serverless multimodal streaming

We use Apache Flink SQL on Kinesis Data Analytics / Databricks Serverless to normalise player events (Hls.js, Shaka, dash.js) into a common media_event_v2 schema (OpenTelemetry + JSON Schema). Each event is enriched in < 200 ms with:

  • Content: frame/audio embeddings every 5 s (VideoCLIP-XL model quantised to INT8 on vLLM + Ollama over a spot T4 GPU).
  • Context: normalised EPG, rights metadata, mood tags generated by GPT-4o mini (cost < €0.0002/event).
  • Consent: TCF 2.2 hash + automatic purpose mapping (Drools rule).

Real metric: we process 15 TB/day with peaks of 1.2M events/sec; cloud cost < €0.18/TB thanks to spot + Graviton3.

Knowledge layer: hybrid RAG + GraphRAG

Two complementary indexes:

  • Vector (Qdrant / pgvector): chunks of scripts, synopses, reviews, ASR transcriptions (Whisper large-v3) → semantic retrieval for “which scenes generate the most engagement among women aged 25-34?”
  • Graph (Neo4j + GraphRAG): entities (characters, locations, brands), relationships (appears_in, mentions, competes_with) → multi-hop reasoning for brand-lift attribution.

The retriever agent (LangGraph state machine) dynamically decides which index to query and fuses results with Reciprocal Rank Fusion (k=60).

Activation layer: specialised MCP agents

We expose 12 data products via MCP servers (Python + FastMCP) so agents can consume them without coupling: the end goal is closing the productivity gap with AI: decisions in milliseconds.

  • audience_segment_builder: generates lookalike cohorts with CTGAN (differentially private synthetic data, ε=1.2).
  • attribution_simulator: Bayesian counterfactual (PyMC) that estimates incremental lift per creative channel.
  • creative_optimizer: invokes Claude 4 Sonnet (via Anthropic API) to rewrite copy/CTAs according to the content’s mood in real time.
  • yield_forecaster: Prophet + XGBoost over fill-rate, eCPM and floor-price series → dynamic floor recommendation every 15 min.

The agents are orchestrated in CrewAI with human-in-the-loop for decisions over €50k (approval via Slack/Teams using Block Kit).

Anonymised real cases (audited metrics)

Case A: Hybrid national broadcaster (linear + FAST + SVOD)

Challenge: Unify cross-device measurement without persistent IDs; reduce waste in brand campaigns (average frequency 14+).

Solution: Deployment of audience_segment_builder + attribution_simulator in 6 weeks. A probabilistic attribution model based on statistical fingerprinting (IP/UA + timing + content affinity) + clean rooms (AWS Clean Rooms / Snowflake Data Clean Room) with advertisers.

Results at 6 months:

  • 40% reduction in time to deploy new attribution models (from 8 weeks to 3.2 days).
  • ROI of 3.2x in programmatic inventory optimisation (eCPM +28%, fill rate +15%).
  • Average frequency controlled at 3.8 exposures/user/week (−73% waste).
  • 92% accuracy in SVOD churn prediction (AUC 0.94) by integrating FAST + linear engagement signals.

Case B: Sports-focused FAST platform

Challenge: Monetise audience peaks (finals, derbies) where programmatic demand exceeds supply 5x; decision latency < 100 ms.

Solution: yield_forecaster + creative_optimizer agents running on Cloudflare Workers AI (edge) with quantised ONNX models. Dynamic floor price + insertion of contextual creatives generated by GPT-4o (prompt engineering with brand-guideline few-shots).

Results at 4 months:

  • 41% increase in peak-hour revenue (vs. static floor).
  • CTR of contextual creatives 2.3x higher than generic ones.
  • p99 decision latency: 67 ms (target < 100 ms).

Case C: Premium content producer (drama + documentary)

Challenge: Deciding project greenlights at development stage using real audience data + cultural trends, not just “gut feeling”.

Solution: Pipeline of GraphRAG + analyst agent (LangGraph) that ingests: scripts (PDF), focus-group transcriptions (Whisper), historical consumption data (Snowflake), social listening (Brandwatch API), TikTok/Reels trends (Apify). It generates a 3-page memo with: audience sizing, comparable titles, risk factors, marketing hooks.

Results at 12 months:

  • 5 projects greenlit with score > 0.82; 3 already in production.
  • 60% reduction in analysis time (from 4 weeks to 6 days).
  • Estimated savings of €2.1M by avoiding 2 projects with score < 0.35.

The 2025 tech stack: what we really use in production

Layer Technology Version / Model Justification
Agent orchestration LangGraph + CrewAI 0.2.x / 0.80+ Stateful, streaming, native human-in-the-loop
LLM reasoning GPT-4o / Claude 4 Sonnet 2025-08 / 2025-06 Reliable tool use, 128k ctx, optimal cost/quality
Embeddings jina-embeddings-v3 / nomic-embed

Prefer to talk to us directly? Contact the BAOSS team and tell us about your challenge.