This blog post is a brief summary of the ML Summit hosted by Dataiku on July 17th 2025, at Singapore.

Summit Notes

Welcome Address

Speaker: Randy Goh, Area VP, Dataiku

Key Points

The welcome address mainly spoke about the various features of Dataiku present ones and the ones in the roadmap
Dataiku has 800 firms as its customers across the world
Velvet Sundown - 100% AI generated band that shows the importance of AI generated content in the current world

Overall Vision and Keynote: Agents Beyond Chaos

Speaker: Sophie Dionnet, SVP, Product & Business Solutions, Dataiku

Key Points

We are just scratching the surface of:
- True AI transformation
- Real AI analytics
- AI deployment
AI agents are everywhere:
- News from consulting firms
- Cloud providers
- LLM vendors
Major categories:
- Agentic capabilities from AI infrastructure providers:
  - Power
  - Expertise
  - Maintenance
- New startups
- Agents and agent builders in enterprise applications:
  - Connected
  - Siloed
  - Extensible
Enterprise Applications (EAs) are evolving:
- From business rules + databases + transactions ? AI assistants
- Still fit for purpose
- But raise concerns:
  - Cost
  - New silos
  - Relevance for your ecosystem
Cross-application assistants:
- Leading enterprises are longing for a new class of AI-first applications
AI-Extended Data Model
AI Process Automation:
- Example: Automating a marketing content creation process, including legal review
The enterprise AI stack is growing more sophisticated to meet that need:
- Cross-application assistants
- AI-extended data model
- AI-powered processes
Key components:
- LLM Ops
- AI Governance
- AI Agent Builder
- Analytics:
  - Business Intelligence
  - Last-mile ETL
- Models:
  - ML Ops
  - Data Science / ML Platform
- Agents
Universal AI Platform:
- Multi-cloud
- Multi-platform
- Multi-model
- Unified governance
- Agents:
  - Agent creation
  - Agent control
  - Catalysts for going deeper
- Models:
  - Model development
  - Model operations
- Analytics:
  - Data preparation
  - Insights
Agents introduce a new type of intelligence to the enterprise:
- Tasks:
  - Agents become the primary UI
- Roles:
  - Machine-like tasks are automated
Even simple agents are difficult:
- Limited functionality
- Inaccurate data
- Lack of robustness
- Hallucinations
- No way to monitor ongoing quality of outputs
Multi-agent systems have become a chaotic mess:
- Mistakes compound in cascades of errors
- Agents created in siloed systems have become ungovernable
Three common problems:
- Fragile connections to systems
- Lack of accountability
- Impossible to diagnose failures
Now is the time to get agentic architecture and governance right:
- Enterprise orchestration
- Fully governed creation
- Continuous optimization
Dataiku orchestrates:
- Analytics
- Models
- Agents

[Panel] AI in the Force

Moderator: Sophie Dionnet Panelists:

Kian Boon Lim (HTX)
Daniel Seet (SCDF)
Sean Tan (MHA)

Key Points

Paramedics deal with around 40 different protocols for various cases.
- AI is being used to:
  - Analyze case feedback
  - Prioritize the relevant protocols based on case patterns
Robotics
Real-time incident management
Chunk the AI project

Agentic AI: Enterprise Automation, Safety, and Scalability

Speaker: Dr. Clifton Phua, IMDA

Key Points

Agentic AI marks a shift:
- From task-based human work
- To goal-driven AI agents
National AI Strategy:
- Singapore’s strategy focuses on:
  - Operationalizing AI
  - Scaling real-world AI systems
Virtual agents for citizen services:
- Customer service AI agents
- Merchant assistant agents
Project Vend:
- Autonomous office shop pilot (June 2025)
  - Demand response
  - Excessive generosity
  - AI hallucinations
- Resulted in a 20% value loss in one month
- Ultimately deemed a failure
- Demonstrated how safe experiments can reveal important AI failure modes
Implementation challenges:
- Technical limitations:
  - Complex reasoning and long-term memory are still hard for AI agents
  - ~20% behind human-level performance
- System integration:
  - Difficulty connecting AI agents to existing APIs and legacy systems
- Trust and safety:
  - Hard to ensure agent reliability
- Deployment is difficult:
  - Requires practical constraints and guardrails
- ROAI (Return on AI Investment) is unclear
Emerging solutions:
- Technical limitations:
  - Specialized agents for domain tasks
  - Multi-agent architectures
- System integration:
  - Open frameworks
  - Standardized agent communication protocols
- Trust and safety:
  - Embedded guardrails
  - Sandbox testing environments
Recommendations:
- Start with small pilots
- Co-develop solutions with domain experts
System-level risks:
- Goal misalignment
- Conflicting decisions
- Potential data breaches
- Solutions:
  - Shared policy layers
  - System-wide observability
  - Collaborative innovation
Speaker note:
- Dr. Phua created the entire slide deck using an AI agent

Technical Keynote: Multi Agent Orchestration

Speaker: Jed Dougherty, VP Platform Strategy, Dataiku

Key Points

Provides orchestration for multi-agent frameworks at various scales:
- Team scale
- Department scale
- Enterprise scale
Roadmap:
- Agent orchestration for reliable, composite AI
- Analytics embedded across the enterprise
- Operationalized AI governance
EU AI Act readiness

[Fireside Chat] Order in Chaos: Bringing AI to Fruition

Moderator: Grant Case, Dataiku Speaker: Dr. David R. Hardoon, Standard Chartered Bank

Key Points

Key question: How do we increase AI propagation?
Single most critical safeguard for agentic AI:
- Need for simplification
- Example: If “country” is missing in a transaction, can you build an agent that handles just that?
What is meaningful oversight?
- Start with simple, repetitive tasks and apply agentic AI
- Consider augmenting human capabilities — what more can we do?
Most undervalued competency:
- Start with the “known knowns”
12 to 18 months outlook:
- Hyper-personalized experiences are becoming ubiquitous
- Emergence of “invisible banking”

AI That Makes Sense: Sensemaking Systems 2.0

Speaker: Tan Boon Leong, ST Engineering

Key Points

Organization: ST Engineering
Four foundational components:
- Use case
- Agile process
- Tools
- Skills
Timeline:
- 2024 – Agentic AI: Text to Action
- 2025 – Embodied AI
Focus: Human and machine collaboration
- Virtual assistant
- Co-pilot
- Digital mentor
Shift from task automation to decision augmentation
Sensemaking: Going beyond dashboards
- Multimodal data ingestion
- Contextual grounding
- Event intelligence
- Explainable recommendations
Use cases:
- Pattern of Life Analysis:
  - Identify non-compliant vessels operating near or within Singapore waters
  - Maritime security – detect suspicious behavior
  - Uses deep learning (sequence-to-sequence models)
- MRO Product DNA:
  - How to digitalize maintenance and repair operations
- Knowledge graph construction:
  - Often hard to start
  - Use GenAI to:
    - Extract relationships
    - Represent them in graph form
    - Enable natural language querying
Reflections:
- Value creation by enhancing current systems to the next level
- Importance of upskilling

Building Trust in AI

Speaker: Shameek Kundu, AI Verify Foundation

Key Points

AI Verify Foundation (AIVF) is a subsidiary of IMDA
Message: Make AI boring again
Reference: Spain electricity catastrophe
Guidelines are great, but:
- It’s essential to ensure AI is reliable on the ground
GenAI testing:
- Much focus today is on frontier risk
- Critical to ensure models are not misaligned
Shift in focus:
- From safety to reliability
Models must be tested as part of end-to-end systems
Lessons learned:
- 17 companies participated with real-life GenAI applications
- Paired with professional testers to evaluate effectiveness
AIVF’s mission: Make AI model testing reliable
Global AI Assurance Pilot:
- Develop testing norms and practices
- Establish foundations for a viable assurance market
- Contribute to roadmaps for AI testing tools

The Agentic AI Revolution: Transforming Business Vision into Tangible Value

Speaker: Catherine Ko, Amazon Q GTM Lead – APJ, AWS

Key Points

The Age of AI:
- Pre-2020s: Innovation cycles took 18–24 months
- GenAI era: Innovation cycles have accelerated to 4–6 months
- Agents are now capable of performing longer tasks autonomously
- Amazon is investing $100B in AI in 2025
Evolution into Agentic AI:
- Generative AI:
  - Follows a set of rules
  - Automates repetitive tasks
- Generative AI Agents:
  - Achieve singular goals
  - Handle broader ranges of tasks
  - Automate entire workflows
- Agentic AI Tasks:
  - Fully autonomous
  - Built on multi-agent systems
  - Mimic human logic and reasoning
Businesses are creating value with AI agents:
- Workplace productivity
- Business workflows
- Innovation and research
Example: NFL
- Agents accelerate content production
Agents are transforming Amazon’s internal systems:
- 10,000 apps migrated
- 4,500 developer years saved
- $260M in savings
Example: Cognizant
- AI agents used to streamline mortgage workflows
Example: Moody’s
- AI application built to generate comprehensive risk reports
- Time reduced from 1 week to 1 hour
Amazon Q:
- Industry-leading specialized agents
- Amazon Q Developer:
  - AWS AI-powered assistant for software development
AWS Transform Initiatives:
- Java migration
- .NET modernization
- COBOL to Java
Agent Demo:
- Agentic IDE released at Kiro.dev
- Features:
  - Generative AI assistant
  - Brainstorming support
Agents can automate complex workflows
Key takeaway: Put together the building blocks for agent-based automation
Amazon Q Index in Asana:
- Asana integrates Q Index to pull data from various applications
“Buy for Me” feature:
- Rolled out using Amazon AI
Future interfaces:
- Designed to support data scraping by AI agents

Operationalising Data Quality for Agentic AI with Dataiku

Speaker: Sophie Dionnet

Key Points

The agentic world has an Achilles’ heel: Data Quality
Trusted data is the cornerstone of agent efficiency
There is increasing demand for:
- Less intermediated access to data
Trusted data depends on:
- Data quality
- Transparency:
  - Understanding the lineage of the data
- Documentation
- Ownership
Virtuous circle for building data trust:
- Define your rules
- Monitor dashboards
- Troubleshoot issues
- Repeat
Any outliers are sent back (feedback loop)
Column-level data lineage:
- Recently added feature (a few months ago)
Issue resolution:
- Troubleshoot any form of data issues
Documentation practices:
- Exists at all levels:
  - Collection
  - Project
  - Dataset
  - Column
- Uses:
  - Expansible wikis
  - Discoverable knowledge
Metadata generation:
- Use LLMs to assist in generating metadata
Evolving data roles:
- From Data Stewards to Data Influencers
  - Share and reuse golden datasets
  - Publication workflows
  - Secured sharing mechanisms
Dataiku provides tools to operationalize and scale data quality efforts

[Panel] The Future of Enterprise AI Infrastructure

Moderator: Grant Case Panelists:

Geoff Soon (Mistral AI)
Ettikan K Karuppiah (NVIDIA)

Key Points

FSI Agent AI Blueprint:
- NVIDIA builds blueprints for AI transformation
- Today’s AI-native companies are born in the generative AI era
- These are transformation-focused companies
NVIDIA’s position:
- Horizontal player in the AI ecosystem
- Builds:
  - System libraries
  - Application-specific libraries
- Uses microservices as building blocks
- Blueprints are composed over these microservices
- Vertical-specific template applications help FSI build AI applications
LLMs:
- Train models with data ? output parameters (e.g., 7B model)
- Open-weight models:
  - Make weights accessible to customers
  - Enables auditing and control
- Pretrain models to fit specific domains
Agents:
- Designed for end-to-end transparency
Mistral AI:
- Advocates for open-source and open-weight models
What makes NVIDIA’s blueprint different:
- Enterprise AI factory:
  - Tests and validates hardware
  - Software is tested and validated
  - Includes:
    - Frameworks
    - Applications for FSI models
    - Hardened models for specific environments
- NVIDIA microservices:
  - Hundreds of models
  - 400+ libraries
  - Can be composed into complete applications
  - Includes:
    - Pre-processing libraries
    - Post-processing libraries
    - Retriever microservice
    - Reranker microservice
    - Reasoning engine ready for deployment
Mistral AI on infrastructure:
- Customer of NVIDIA
- On-prem deployment has become cost-effective
Dataiku capabilities:
- Helps manage prompts to ensure governance
- Supports LLM mesh: choose any LLM you want
- Empowers individuals with:
  - Access to data
  - Capabilities to integrate data and agents
- Enables tighter integration between enterprise data and agents
Blueprint implementation:
- Dataiku serves as an orchestration platform
- Manages data access and control
- Simplifies workflow implementation when driving use cases
Example use case: AML (Anti-Money Laundering)
- Use the right datasets
- Data governance essential
- Use NIMS:
  - Guardrails at input and output
- Retrain LLMs using enterprise datasets
- Trace and track communication between agents:
  - Important for explainability
Open-weight LLMs:
- The gap between public cloud and on-prem has narrowed
- Shift from personal productivity to enterprise transformation
- Focus on:
  - Specific models
  - Strong governance
  - Moving from human interface to agent interface
For highly regulated industries:
- Work closely with data owners
  - Understand what data and applications can be used in agentic AI
  - Identify low-hanging fruit for automation
- Collaborate with LLM providers
- Do a quick PoC
- Keep human-in-the-loop as part of oversight

[Panel] From Promise to Practice: AI, GenAI & Agents

Moderator: TC Gan Panelists:

Sandeep Moturi (MSD)
Melvyn Peh (STMicroelectronics)
Ann Tey (NTU)

Key Points

Bio/Life Sciences Vertical:
- Current processes involve tests like angiograms
- New techniques such as digital twin:
  - Capture a signature of your heart
  - Predict long-term heart function using AI
- GenAI accelerates classical AI:
  - Awareness ? Diagnosis ? Treatment
  - Apply AI on digital twin simulations
- Use case from South Korea:
  - Boosted sales by 10%
  - Purchased property data, mapped lat/long of Seoul
  - Identified that 20% of potential customers were not being targeted
  - No AI involved – pure business analytics
- Use case with AI for pricing injections:
  - Inputs:
    - Social media sentiment
    - Past sales data
    - Doctor sentiment
  - Create a model to:
    - Maximize sales
    - Minimize reputational risk
    - Visually identify the pricing “sweet spot”
- GANs used for image generation:
  - Competing models generate optimal visual outputs
High Tech Manufacturing:
- In the semiconductor industry:
  - Data has always been present
  - Each product passes through hundreds of machines
  - AI makes machine interactions intelligent
  - Helps engineers predict and respond proactively
- Legacy challenge:
  - Data was siloed
  - Now with Dataiku, systems can integrate and interact
- Data availability is no longer the bottleneck
- Chatbots:
  - Support users by accessing the knowledge base
- Vision:
  - Move towards dark factories
  - Employ humanoids for autonomous operations
- NVIDIA provides tools for synthetic data creation
- Physical AI:
  - The bleeding edge
  - Use LLMs to generate training data for robots
Higher Learning:
- Use cases across admin, faculty, and research:
  - Connect data across silos
  - Use it for machine learning applications
- Example:
  - Identify students at risk of failing
  - GenAI used to respond to helpdesk queries
- Physical AI in research:
  - Example: solar panel chip design
  - Use LLMs to instruct robots to create advanced chips
- NTU scale:
  - 30,000 students and 3,000 staff
  - Focus on democratizing AI within the university
- Tools:
  - Use AI Co-pilot tools to democratize access and usability
AI Security & Governance:
- Emerging role: AI Guardian
  - Similar to a security guard or police department for AI
- Need for a governing body or MCP (Model Control Panel)
  - Responsible for oversight, policy, and standards
Closing Thought:
- The entire AI ecosystem needs to grow together

Idea to Impact: How Dataiku Powers End-to-End AI

Speakers:

TC Gan
Ming Zuo Liew

Key Points

Overview:
- Agents are everywhere
- Enabling a new class of enterprise applications
- But building those applications can be chaotic
Market trend:
- 25% of firms doing GenAI will deploy an agent in 2025
- That number is expected to reach 50% in two years
Universal AI Platform:
- Designed for agent creation and control
- As applications and agents multiply, things get messy:
  - Multiple teams building multiple agents
  - Leads to complexity and technical debt
Demo theme:
- Not just about innovation, but about rapid creation
- AI project lifecycle:
  - POC ? Pilot ? Product ? Critical Product
Live Demo:
- Example: Singapore Airlines
  - From data ingestion to dashboard insights
  - Steps:
    - Ingest XML data
    - Prepare using recipes
    - Use GenAI for:
      - Entity extraction
      - Dashboard creation
    - Ask questions on top of the results
Enterprise Orchestration:
- Coordinate multi-agent and multi-model setups with security and guardrails
- Components:
  - LLM Mesh:
    - Securely manage connections to multiple LLMs
    - Configure specific LLMs for specific use cases
    - Centrally manage all LLM connections
  - Sage Guard:
    - Control panel for risk
    - Detect PII in prompts
    - Detect prompt injection attempts
  - Agent Connect: Connect and manage agents
  - Cost Guard: Monitor and optimize usage costs
  - Quality Guard: Ensure output quality
  - Trace Explorer: Observe agent behavior and lineage
Analogy: Think of an F1 race setup
- Each car = an AI Agent
- F1 Pit Lane = LLM Mesh infrastructure
  - Where agents go for updates and orchestration
- Sage Guard = Safety crew
- Cost Guard = Race strategist
- Quality Guard = Performance analyst