The value of an integrated Big Data platform

June 3, 2025 · Baoss
BAOSS IT consulting: unified data platforms that turn the data lake into an AI agent engine

BAOSS IT consulting: unified data platforms in 2025, from data lake to AI agent engine

In June 2015 we published on this blog that “the value of the integrated Big Data platform is its capacity for analyzing structured and unstructured data at the pace companies need”. Ten years later, however, that sentence sounds almost naïve. The pace is no longer set by the nightly batch or Apache Kafka streaming. Now it is set by autonomous agents that reason, call tools and make decisions in milliseconds. At BAOSS IT consulting we see this every week with our clients. If you wonder whether you are ready, read AI agents: is your company exposed to the shock of disruption?.

The real problem in 2025-2026 is not storing petabytes, because object storage costs pennies. It is not even processing them. The problem is contextualizing the information. The goal is for an LLM (GPT-4o, Claude 4, Llama 3.1 405B) to be able to act on it without hallucinating. And that is where the old “integrated platform” mutates into a data platform ready for AI agents. This includes a unified catalog, automated governance and column-to-prompt lineage. It also includes a semantic layer that understands the business, not just schemas.

The invisible bottleneck: context, not volume

Take two references: Databricks’ State of Data + AI 2024 and a16z’s Data & AI Landscape 2025. According to both, 78% of Spanish organizations that started GenAI projects in 2023 have stopped or slowed down their production deployment. The main cause is not the model, because GPT-4o and Claude 4 Sonnet already handle complex reasoning tasks. The cause is the absence of a data layer. It must be reliable, versioned and accessible via function calling or MCP (Model Context Protocol).

  • Fragmentation of sources: for example, mainframe (COBOL/VSAM), ERP (SAP S/4HANA, Oracle Cloud), SaaS (Salesforce, Workday), industrial IoT, Kubernetes logs, scanned PDFs, emails, contracts.
  • Broken semantics: “customer” in CRM ≠ “counterparty” in core banking ≠ “buyer” in e-commerce. Without a unified semantic layer, therefore, the agent confuses entities and hallucinates relationships.
  • Reactive governance: access policies, PII, lineage and quality are applied ex post and, therefore, not at inference time.
  • Runaway inference cost: naïve RAG injects 50k tokens per query. However, prompt caching and semantic routing cut API spend by 60-70%.

At BAOSS IT consulting we have measured this gap across twelve enterprise accounts (banking, retail, logistics, energy) during 2024-2025. The finding: 65% of a GenAI project’s time is consumed by “pre-LLM” data engineering —cleansing, reconciliation, feature store, groundedness evaluation—. That is why the modern unified platform exists: to compress that 65% down to under 20%, that is, to less than a third of today’s effort. And when it is time to scale, these criteria for choosing an AI technology partner in 2026 will serve as your compass.

Reference architecture 2025: layers, not silos

First, the stack we deploy today at clients —on AWS, Azure or GCP, hybrid or air-gapped— follows this logical model:

  • Unified ingestion: Debezium CDC + Kafka Connect + Airbyte Cloud for SaaS; native zero-ETL (Aurora → Redshift, Spanner → BigQuery). End-to-end latency < 2 s.
  • Open lakehouse: next, Apache Iceberg / Delta Lake / Apache Hudi on S3 / ADLS / GCS. Single table format, time travel, ACID, schema evolution without rewrites.
  • Semantic & metrics layer (Headless BI): Cube.dev, dbt Semantic Layer or AtScale. They define “net revenue”, “churn”, “available stock” once; consumed by Looker, Tableau, and by agents via MCP/Function Calling.
  • Hybrid Feature Store & Vector Store: Feast + pgvector / Qdrant / Pinecone / Milvus. Embeddings generated with text-embedding-3-large or bge-m3 (local Ollama for sensitive data). Semantic chunking, not fixed-token chunking.
  • Agent orchestration: LangGraph (stateful, cycles, human-in-the-loop), CrewAI (collaborative roles), AutoGen (multi-agent conversation). Deployed as FastAPI microservices + vLLM / TGI for open models (Llama 3.1, Nemotron 3 Ultra, Qwen 2.5) or managed endpoints (Azure OpenAI, Bedrock, Vertex AI).
  • Observability & Guardrails: LangSmith / Langfuse / Arize Phoenix for traces, online evaluation (faithfulness, answer relevance, PII leakage), prompt canaries, human fallback.
  • Native governance: finally, Unity Catalog / OpenLineage / DataHub. Attribute-based (ABAC) policies that the agent respects at runtime via a policy enforcement point (OPA / Cedar). See our lesson on AI agent governance (the Hugging Face case).

In short, the key is that every layer exposes standard APIs (REST, gRPC, MCP, SQL over HTTP). That way, the agent —or the human analyst— consumes data without coupling to the underlying technology.

Case 1: Wholesale banking — Fraud detection in instant SEPA payments

Context: top-5 Spanish financial institution, 12 M customers, 3,500 M transactions/year. In other words, a very high volume. EBA/PSD2 regulation requires a decision in < 10 s on instant transfers. Legacy system: Drools rules + an XGBoost model retrained monthly, 2.3% false positives, 0.8% undetected fraud.

BAOSS solution (6 months, team of 8):

  • Migration to an Iceberg lakehouse on S3 + Trino for federation of mainframe (VSAM) + core banking (Oracle Exadata) + digital channel (Kafka).
  • Cube.dev semantic layer: “transaction”, “originator”, “beneficiary”, “device”, “location” entities with hierarchies and enriched attributes (IP reputation, device fingerprint, beneficiary graph).
  • FraudReasoner agent (LangGraph + GPT-4o mini fine-tuned on 120k labeled cases): receives an ISO 20022 event, queries the feature store (Feast + Redis), retrieves vector k-nearest neighbors (similar historical patterns) and runs chain-of-thought reasoning. It then returns APPROVE / REVIEW / BLOCK with a natural-language explanation for the analyst.
  • Guardrails: P99 latency < 8 s and fallback to a lightweight model (Llama 3.1 8B on vLLM) if the external API fails. In addition, full auditing in DataHub.

Results after 9 months in production:

  • False positives: -42% (1.33%).
  • Undetected fraud: -61% (0.31%).
  • Average decision time: 3.7 s (P99 7.2 s).
  • Direct savings from avoided fraud: €18.4 M/year.
  • Project ROI: 4.1x in 12 months.
  • Reduction in feature engineering time for a new pattern: from 3 weeks to 2 days (semantic layer + self-documenting agent).

Case 2: Omnichannel retail — Conversational shopping assistant with real inventory

Context: fashion/homeware retail group, 1,200 stores + e-commerce, 45 M active SKUs, 18 M loyalty customers. In other words, a complex omnichannel operation. Goal: a WhatsApp/Web assistant that reserves stock in < 5 s. For example, it must answer "do you have this t-shirt in size M at a nearby store?".

BAOSS solution (4 months, team of 6):

  • First, data unification: SAP S/4HANA (stock), Salesforce Commerce Cloud (catalog), CDP (customer profile), Google Analytics 4 (browsing). Iceberg + dbt for incremental transformations every 15 min.
  • Semantic layer: “real availability” = physical stock – pending reservations – online orders not picked up + transfers in transit. Computed as an incremental materialized dbt model and then exposed via the Cube.dev SQL API.
  • ShopAssist agent (CrewAI: Router → InventoryChecker → Recommender → Responder). InventoryChecker uses function calling over an MCP server that wraps Cube.dev; Recommender retrieves product embeddings (bge-m3 in on-prem Ollama) + customer history to suggest alternatives when there is no stock.
  • Hybrid RAG: product catalog (vector) + return/promotion policies (Neo4j knowledge graph + Cypher generated by the agent).
  • Continuous evaluation: Langfuse + dataset of 5k labeled real conversations; groundedness > 0.92, hallucination rate < 0.8%.

Results after 6 months:

  • Level-1 resolution (no human): 78% (previously 34% with a rules-based chatbot).
  • Assisted conversion: +23% in sessions with the agent vs. control.
  • Chat channel NPS: +18 points.
  • Cost per conversation: €0.042 (GPT-4o mini + semantic caching) vs. €0.38 for a human agent.
  • New product category deployment time: -85% (weeks → hours, just update catalog + embeddings).

Case 3: Energy — Multi-asset predictive maintenance with a lightweight digital twin

Context: renewables operator (wind + solar + batteries), 2.4 GW, 1,100 wind turbines, 3,200 inverters, 450 BESS. Data: SCADA (1 Hz), CMS (vibration, 10 kHz), drone inspections (images), OEM manuals (PDF), SAP PM tickets. The goal is twofold: anticipate critical failures 14-21 days ahead. In addition, optimize scheduled shutdowns.

BAOSS solution (8 months, team of 10):

  • First, ingestion: Kafka Connect + OPC UA for SCADA/CMS; Airbyte for SAP PM; image pipeline → YOLOv8 (blade defects) → Iceberg metadata.
  • Lakehouse: Iceberg partitioned by asset_id / year / month; time travel for training reproducibility.
  • Feature store (Feast): 340 features per asset (statistical, spectral, oil degradation, ERA5 weather, maintenance history).
  • AssetGuardian agent (AutoGen: DataAnalyst + ReliabilityEngineer + Planner). DataAnalyst queries the feature store + vector store (OEM manuals with semantic chunking + Llama 3.1 70B embeddings). ReliabilityEngineer reasons with structured chain-of-thought (JSON schema) and generates: probabilidad_fallo_14d, modo_fallo_top3, acciones_recomendadas, confianza. Planner integrates with SAP PM to create a preventive work order if confianza > 0,82.
  • Base models: XGBoost + TabPFN (tabular) + 1D CNN (vibration) ensembled; in addition, the agent decides the weight per asset. Monthly retraining is automated (Airflow + MLflow).
  • Governance: Unity Catalog + column-to-prompt lineage; ABAC policies per role (operator, engineer, management).

Results after 10 months:

  • Unplanned critical failures: -57%.
  • Unplanned downtime hours

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