Session

AskPorto — a Unified Agentic AI Solution for Data Analytics and Discovery

Overview

Experience	In Person
Track	Artificial Intelligence & Agents
Industry	Public Sector
Technologies	Databricks SQL, Unity Catalog, Genie
Skill Level	Intermediate
DOWNLOAD SESSION SLIDES

The Central Provident Fund Board (CPFB), Singapore’s social security agency, manages complex data across multiple business lines with unique analytical requirements despite centralising on Databricks Lakehouse. Users struggled with traditional querying methods and fragmented governance documentation across multiple sources, creating analytical silos. Employees needed intelligent systems capable of querying multiple sources without technical expertise, requiring contextualised responses understanding domain-specific terminology and automatic validation. Solution: unified conversational interface with multi-agent orchestration built on Databricks architecture. Orchestrating AI routes queries to specialised agents — Databricks Genie for contextualised data answers, Purview for metadata discovery, RAG-powered agents for governance knowledge — whilst maintaining audit trails and compliance. Stakeholders access comprehensive insights via conversational queries with automated accuracy checks.

Session Speakers

Alex Chen

/Senior Consultant
Central Provident Fund Board Singapore

Benedict Ho

/Senior Deputy Director
Central Provident Fund Board Singapore

Jie Ning Ng

/Lead Analyst
Central Provident Fund Board Singapore

Full Summary

Inside Singapore CPF Board's AskPortal: Building Agentic AI on a Government-Grade Data Platform

Singapore's Central Provident Fund (CPF) Board spent more than a decade modernizing its data estate, evolving from mainframe-era silos to a unified, cloud-based foundation that now supports predictive models and generative AI.

The session introduces AskPortal, an agentic AI assistant that lets non-technical staff query policy documents and governed datasets in natural language, illustrating how a highly regulated agency can adopt conversational analytics without loosening control — a challenge explored in depth by practitioners weighing governed self-service analytics against existing BI investments.

A governance-first approach underpins the effort. The team shows how strong identity, metadata, and platform architecture enable practical AI outcomes, shifting access from raw tables to on-demand insights while preserving trust.

From fragmented systems to a unified, governed foundation

CPF's path ran from early systems like SIB and the DANIEL analytics lab to the Unified Data Platform (UDP) on Microsoft Azure and the Databricks Platform. The goal was a one-stop environment that scales from simple reads to generative AI, supports standardized MLOps, and avoids lock-in.

Data enters via Azure Data Factory into Databricks-managed bronze, silver, and gold layers. Sensitive fields are tokenized or re-encrypted on landing, then decrypted dynamically through secure views for authorized users. Access is governed by Azure Entra ID groups synced to Databricks, and data discovery is centralized in Azure Purview. Workloads integrate with Power BI for dashboards, Neo4j for graph use cases, and Azure Kubernetes Service for model hosting.

Operational wins that validated the platform

Before layering agentic AI, UDP delivered measurable improvements:

A contact center forecasting model simulates demand scenarios across more than half a million annual calls, enabling proactive staffing and shorter wait times.
A real-time classifier routes written inquiries to the correct department the moment they arrive, cutting misrouted cases by about one third and speeding responses.
A personalized learning solution combines UDP, a Neo4j knowledge graph, and an LLM-powered chatbot so employees can ask "what's next for me?" and receive tailored learning recommendations by role, skills, and trajectory.

AskPortal: agentic AI that understands both policy and data

AskPortal removes two barriers to self-service insights, domain context and coding skills. A main agent routes each question to a wiki agent for policy and process queries, powered by a vector store of internal documents, or to a data agent that executes queries through Genie spaces mapped to specific topics, such as member demographics. Conversation context carries across follow-ups, allowing users to refine answers naturally.

Security design separates the chat interface, orchestration engine, and data context across distinct virtual networks. Azure Kubernetes Service hosts application pods, while Genie integrates natively with Unity Catalog and Databricks SQL. User identity is preserved end to end via an on-behalf-of token swap, exchanging the front-end token for a Databricks-scoped token that Unity Catalog validates, so the AI sees only data the user is authorized to access.

What made Genie reliable: metadata and methodical tuning

Metadata quality proved decisive. The team rewrote cryptic table and column names into plain business language and improved descriptions to reduce ambiguity and hallucinations. Genie was taught internal acronyms, explicit join paths, and reusable SQL expressions that bind business terms to logic.

Example end-to-end SQL queries for common questions outperformed long textual instructions. Development followed a tight build, measure, and optimize loop, starting with single-table queries, then introducing joins. A static test suite of golden questions with verified SQL enabled consistent benchmarking and isolated change testing.

Governance-led design turned AI into a daily capability

Agentic analytics did not hinge on model choice, it hinged on disciplined identity, metadata, and architecture. By baking governance into every layer, CPF Board made conversational access to sensitive, trusted data routine for business users, without diluting controls that keep the data safe.

FAQ

AskPortal is CPF Board's internal agentic AI assistant that answers natural language questions about internal policies, processes, and governed datasets. It is designed for business users who lack SQL or Python skills, enabling true self-service insights.

User identity flows through an on-behalf-of token swap. The front-end token is exchanged for a Databricks-scoped token that Unity Catalog validates, so the AI can only query data the specific user is authorized to access under existing role-based controls.

Genie provides a built-in knowledge store for synonyms, column descriptions, and example SQL, which improves translation from English to queries. It also inherits Unity Catalog governance natively, an essential requirement given varying data sensitivity.

Clear, business-friendly metadata, explicit join mappings, and reusable SQL expressions reduced ambiguity. Concrete example SQL for end-to-end questions, plus an incremental build and golden-question benchmarking loop, produced reliable results faster than long text prompts.

A main agent interprets intent and routes to a wiki agent for policy and process questions or to a data agent for statistics. The data agent maps requests to targeted Genie spaces by topic, narrowing scope for more accurate execution.