HungerStation · Internal · 2025 RAG · LLM Production

RAG Analytics
Assistant

A Gemini-powered self-service analytics assistant that lets stakeholders query business data in plain language — backed by an automated BigQuery metadata pipeline spanning 100+ table schemas across HungerStation's Local Shops vertical.

Internal project — architecture and approach shared at a high level. Proprietary data, schema details, and implementation specifics are confidential.
Year
2025
Organisation
HungerStation
Role
Designer & Builder
Status
Production

The problem

HungerStation's Local Shops division runs 3 Looker Studio dashboards spanning 53+ pages of operational, commercial, and financial data. Every week, the analytics team received a backlog of repetitive stakeholder requests — "what's the revenue for chain X this month?", "which branches are underperforming on delivery time?" — questions answerable from existing dashboards, but requiring manual intervention to surface.

The root cause wasn't a data gap. It was a discoverability gap: stakeholders didn't know what data existed, what tables to query, or how to navigate 53 pages of dashboards efficiently.

Architecture

Layer 1 — Automated knowledge base ingestion
BigQuery
100+ table schemas, SQL patterns, relationships
→
Metadata Pipeline
Automated extraction & cataloguing
→
Vector Store
Embeddings of schema + query patterns
Layer 2 — RAG retrieval & generation
Stakeholder Query
Natural language question
→
Retriever
Semantic search over knowledge base
→
Gemini 1.5 Pro
Context-grounded generation
Layer 3 — Output
Answer + SQL
Grounded response with source references
→
Dashboard Pointer
Links to relevant Looker Studio page
→
Stakeholder
Self-served, no analyst needed

Capabilities

◈
Natural language querying
Stakeholders ask questions in plain English or Arabic — the assistant routes to the right schema and generates grounded answers.
⬡
Automated schema cataloguing
A pipeline continuously extracts BigQuery table schemas, column definitions, SQL query patterns, and table relationships into the knowledge base.
◎
Dashboard navigation
Responses include direct pointers to the relevant Looker Studio dashboard page — 53+ pages indexed and searchable by topic.
▦
SQL generation
For custom queries, the assistant generates BigQuery-compatible SQL grounded in the actual schema — no hallucinated columns.

Impact

100+ Table schemas indexed
53+ Dashboard pages made searchable
↓ Repetitive ad-hoc analyst requests

The assistant shifted the analytics team's role from reactive query-answerers to proactive insight-generators — a meaningful operational change for a division handling high-volume transactional data across multiple retail chains.

Built with

Gemini 1.5 Pro BigQuery Python RAG Pipeline Vector Embeddings Looker Studio GCP Airflow
© 2026 Khalid Alnujaidi ← Back to portfolio