Building a Self-Service Data Analytics Platform with Generative AI and RAG

zeb · Tech · 2024 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise self-service data analytics

zeb needed to reduce the manual workload on data analysts by letting business users request reports, forecasts, and insights in natural language through existing communication tools while preserving enterprise governance and industry-specific context.

Constraints:

• Business users already submitted data requests through channels such as Slack, Teams, and email
• Data analysts carried a substantial backlog of manual requests
• The product needed to support logistics, retail, fintech, and health/life-sciences contexts
• The system needed both customer-specific schema grounding and industry-specific model behavior
• Enterprise customers needed governed access to data, model serving, and fine-tuning assets
• Outputs needed to route to warehouses, AutoML endpoints, or reporting tools

Architecture

SuperInsight is a compound AI system built on the Databricks Data Intelligence Platform. End users send requests through email, Slack, Teams, or other channels. A DBRX model classifies intent, Databricks AI Search retrieves knowledge-base context stored in Unity Catalog, and a Model Serving endpoint combines DBRX with a fine-tuned industry adapter. The final output routes to a data warehouse for CSV generation, an AutoML endpoint for forecasts, or a reporting tool for visual reports.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  user-channels[User Channels] --> dbrx-intent-classifier[DBRX Intent Classifier]
  dbrx-intent-classifier[DBRX Intent Classifier] --> databricks-ai-search[Databricks AI Search]
  dbrx-intent-classifier[DBRX Intent Classifier] --> model-serving-endpoint[Model Serving Endpoint]
  databricks-ai-search[Databricks AI Search] --> unity-catalog[Unity Catalog]
  model-serving-endpoint[Model Serving Endpoint] --> unity-catalog[Unity Catalog]
  model-serving-endpoint[Model Serving Endpoint] --> data-warehouse[Data Warehouse]
  model-serving-endpoint[Model Serving Endpoint] --> automl-endpoint[AutoML Endpoint]
  model-serving-endpoint[Model Serving Endpoint] --> reporting-tool[Reporting Tool]

Components

Data flow

A user submits a natural-language data request through a familiar channel. The request goes to a DBRX classifier, then Databricks AI Search retrieves relevant context from Unity Catalog. Model Serving combines another DBRX model with a fine-tuned adapter for the customer's industry. Depending on the intent, SuperInsight sends output to the data warehouse, AutoML forecasting endpoint, or reporting tool.

Why this architecture

zeb used a compound AI architecture because fine-tuning and RAG solved different parts of the analytics problem: fine-tuning adapted model behavior to industry context, while RAG grounded the model in each customer's schema and data context. Keeping the stack inside Databricks consolidated model serving, security, federation, RAG, and training instead of stitching many systems together.

Models used: DBRX · DBRX Mixture-of-Experts · fine-tuned industry adapters · AutoML models

Scale

• zeb has over 15 years of experience and more than 1,500 experts.
  “With over 15 years of experience and a team of more than 1,500 experts, zeb is a leader in digital transformation strategy.”article, stated
• The initial deployment reduced analyst workload by up to 80-90%.
  “processing data requests, the system transformed user requests into actionable insights, significantly reducing the workload on data analysts by up to 80–90%. It seamlessly integrated with the customer’s existing workflow, such as Slack channels, where data requests were already managed. The”article, stated
• zeb targets four industry families with canonical data models.
  “teams, and at a price point that will give them real value and real results very quickly.” They decided to apply a compound AI system that could leverage internal canonical data models aligned with those four different industries.”article, stated
• Databricks reduced solution development time by 40%.
  “Stated Sid Vivek, “We can confidently say we reduced the time to develop our solution by 40% using the Databricks Platform. The customer’s data is in Databricks.”article, stated
• Customers saw 40% cost savings from augmenting analyst teams.
  “Claimed Sid Vivek, “By using SuperInsight to augment the work of their data analyst teams, customers are seeing a 40% cost savings while their data analysts are able to focus on more critical tasks.” Another beneficial outcome was that more end users began interacting with their data to gain”article, stated
• Report requests increased by 72% compared with the manual process.
  “We actually saw a 72% uptake in the number of reports requested from our generative AI solution vs. the previous manual process”article, stated

Alternatives considered (roads not taken)

Keep manual analyst request processing

Rejected because: Analysts had a substantial backlog, and SuperInsight could turn natural-language requests into actionable insights.

Use only fine-tuning or only RAG

Rejected because: Fine-tuning and RAG addressed different context layers: industry behavior and organization-specific schema/data.

Assemble many separate LLMOps products

Rejected because: Databricks consolidated data, fine-tuning, model serving, security, and federation in one platform.

Patterns

rag-retrieval (retrieval)

SuperInsight retrieves organization-specific context from a Unity Catalog knowledge base.

fine-tuning-adaptation (model adaptation)

zeb uses fine-tuned adapters to adapt DBRX to specific customer industries.

model-routing (platform)

The pipeline classifies intent and routes output to CSV generation, AutoML forecasting, or report generation.

tool-use-orchestration (integration)

Natural language requests are converted into actions across Databricks search, model serving, warehouses, AutoML, and reporting tools.

Tradeoffs

Build a compound AI system

Chose RAG plus fine-tuning on Databricks over A single plain LLM endpoint.

Rationale: The model needed both industry behavior adaptation and customer-specific schema/context grounding.

Use DBRX

Chose Open-source DBRX with Mixture-of-Experts architecture over An unspecified generic model stack.

Rationale: zeb valued instruction fine-tuning, reduced latency, fast inference, and accuracy.

Meet users in existing communication tools

Chose Slack, Teams, email, and other channels over A separate analytics request interface.

Rationale: The customer workflows already lived in these channels, so integrating there reduced adoption friction.

Gotchas (what bit them)

Self-service analytics can increase demand.

Root cause: Making reports easier to request led more end users to interact with data.

Resolution: The platform treated increased request volume as a positive outcome and used the Databricks stack to support the demand.

A unified platform simplifies operations but creates ecosystem dependence.

Root cause: SuperInsight consolidates fine-tuning, serving, security, federation, and data inside Databricks.

Resolution: zeb accepted the consolidation because it reduced implementation complexity and time.

Evaluation

zeb reports workload reduction, development-time savings, customer cost savings, increased report requests, and continued implementation feedback into canonical data models.

Metrics: 80-90% analyst workload reduction · 40% development-time reduction · 40% customer cost savings · 72% increase in report requests · four industry-aligned canonical data model families

Outcomes

• SuperInsight reduced a logistics customer's analyst workload by up to 80-90%
• Databricks reduced zeb's development time by 40%
• Customers saw 40% cost savings from analyst augmentation
• Report requests increased by 72% compared with the previous manual process
• zeb continues to feed implementation learning back into canonical data models

The source describes zeb as a 1,500+ expert consultancy and SuperInsight as a multi-industry product spanning logistics/supply chain, retail, fintech, and health/life sciences.