AISEA Logo
← Events
DBS Research MDirector +AI High Fashion Design (Valentino) +AI Frontier Model Security Defensive Teaming + Gemini Embedding 2 Simplified RAG

DBS Research MDirector +AI High Fashion Design (Valentino) +AI Frontier Model Security Defensive Teaming + Gemini Embedding 2 Simplified RAG

Hosted by AINight

About this event

This is the tentative date. Actual date will be confirmed in June/July.
Welcome back to our meetup in August. The date is right after the August holidays. (If this is your first time coming to the venue, please scroll down to the bottom to find the instructions for getting to the venue.)
06:00 PM Registration, dinner, networking
06:30 PM Talk starts
If you want to have practical AI Agents that can productively search and find relevant rich content for you, this event is for you.
Dinner will be actual food, with real protein like actual beef or chicken in it, not random baloney pizza.

Other community news

  1. We are also planning to hold a social Pickleball event in Jun/Jul/Aug sponsored and open to previous attendees of previous community events. A separate invite and event page will be sent out later.
  2. We may also do something cool for F1 weekend in Oct for previous attendees of at least 3 previous community events.
  3. If you would like to follow our future events, but you are unable to make for this specific event but you would like to attend a future event: Just keep your status as "Invited" without registering and do not select "Not Going" because the Luma platform drops the "Not Going" people from future event invitations. Unfortunately the Luma platform handles it in a way that is counterintuitive to common sense.

Talk 1: AI-Enabled Haute Couture High Fashion Design & Jewellery Design - by Valentino Designer

Talk 2: Protecting your stuff from AI-Powered Attackers (with frontier models like Mythos, Kimi, DeepSeek) that Change the Threat Model
This talk argues that the next major security challenge is survivability under accelerated attack by AI-powered attackers using frontier models like Anthropic Mythos, ChatGPT Pro, or even open weights models like Kimi, DeepSeek Pro.
They change the threat model by compressing every stage of an intrusion. Frontier models can help find vulnerabilities, write exploits, interpret logs, chain weaknesses, and guide lateral movement at a pace that breaks many existing security assumptions.
Once an attacker gains an entry point, the time between initial access and meaningful control can shrink from days to minutes.
Cloud platforms, developer machines, CI/CD systems, package registries, identity providers, and internal tooling all become higher-value targets when AI helps attackers move faster, adapt to unfamiliar systems, and exploit supply-chain trust.
Defensive teams will face a world where patch cycles, manual triage, and broad implicit trust are too slow.

Talk 3: AI in the Financial Center Banking Industry + Intention Economy for Durable Value - by DBS Research Group Managing Director
We will look at the transformative influence of Artificial Intelligence (AI) on financial centers and the banking industry. It introduces a conceptual framework for a Trusted Al Financial Hub, powered by a self-reinforcing Trust Flywheel that balances trust, and adoption.
Introduction to the concept of an "intention economy" in the context of Asia's Al development, arguing that as Al capabilities become commoditized, the primary source of durable value shifts from technological prowess to "intention." This "intention" involves making deliberate choices about which problems to solve, for whom, and under what ethical and governance terms, particularly for high-consequence applications where trust is paramount. It suggests that investors should reallocate capital towards regulated Al, physical infrastructure, climate solutions, and trust-building initiatives, anticipating that such intention-led companies will demonstrate convex returns over a longer duration due to enhanced capital efficiency and revenue durability from accumulated trust. This approach emphasizes building for domestic depth and geopolitical resilience by focusing on locally embedded, high-stakes problems rather than relying solely on generic capability.

PLUS: Q&A for Gemini Embedding 2 Simplified RAG Search (image -> document)
The previous speaker for the Gemini Embedding RAG Search talk is back in Singapore. You can now ask him questions live in person.

Previous Talk

Practical AI from Gemini Embedding 2 for Simplified RAG Search with Rich Context and Less Preprocessing - Imagine the possibilities if you could find a document using an image, or find audio from a report, or find a video with a sentence.
Imagine the possibilities if you could find a document using an image, or find audio from a report, or find a video with a sentence.

  • Simplifies RAG Systems (one multimodal retrieval pipeline instead of multiple fragmented systems)
  • Less preprocessing (e.g. no transcription needed) less infrastructure complexity
  • Richer context (text + visuals + audio)
  • A major step from “document chatbots” to truly context-aware AI systems
    Gemini Embedding 2 changes what Retrieval-Augmented Generation (RAG) can fundamentally do.
    Instead of treating text, images, audio, video, and documents as separate systems, AI can now retrieve and reason across all of them together in one shared understanding layer.
    This means users can search videos with natural language, retrieve diagrams from spoken conversations, match screenshots to documentation, or ground AI agents with real-world multimodal context — capabilities that were previously complex or unreliable without multiple specialized pipelines.

Future Talk

Open Weight Models: What Works, What Breaks, and When to Bet on Them
Cloud APIs have dominated the AI conversation, and for good reason. They're easy to start with. But open weight models have quietly crossed a capability threshold that changes the calculus for businesses that care about cost, data sovereignty, or simply not being locked into someone else's roadmap.
This talk cuts through the hype with a practitioner's lens. We'll start with what actually works in production today: which model families deliver real results on real business tasks, what they handle well, and where they still fall short. Every claim is grounded in actual deployments, not benchmarks or blog posts.
Then we'll get honest about the gotchas. Quantization tradeoffs that silently degrade quality. The gap between "it works in the demo" and "it works at 2am on a Saturday." The operational overhead that nobody puts in the pitch deck. These are lessons learned the hard way, and you should hear them before you commit.
The centrepiece is a live model shootout: open weight models versus cloud APIs, head to head, on tasks that mirror what businesses actually need. Not synthetic benchmarks. Not cherry-picked examples. You'll see where open weight models match or beat the cloud, and where they genuinely don't.
We'll close with a decision framework built from production experience: when open weight is the right bet, when cloud APIs are the better call, and the hidden costs on both sides that most comparisons miss. Whether you're a founder evaluating your AI stack, a CTO weighing build-versus-buy, or an executive trying to separate signal from noise, you'll leave with a practical scorecard you can act on.

Future Talk: TBA

Awaiting confirmation for this Future Talk

Federated Layer to Manage a Multi-Agent Synthetic Workforce
Agents are starting to act as workers in real production systems, and workers need a management layer. The question everyone should take back to your laptop: in my own agent stack, what happens on the second concurrent write, and can I prove, today, under whose authority my agent acted? How do I manage my synthetic workforce in production.

Awaiting confirmation: Talk: Federated Layer to Manage a Multi-Agent Synthetic Workforce
AI is moving from giving answers to taking actions inside real systems. That shift turns agents into a kind of workforce, and a workforce needs management: rules for what each agent can do, memory of what it has already done, coordination across tools, and a durable record of every action. This is a high-level look at why running autonomous agents in production is becoming an infrastructure problem, what tends to break first as agent count grows, and what it takes to make agent execution reliable and accountable. Aimed at builders, engineers, and founders thinking about where agent infrastructure is heading.
Agents are starting to act as workers in real production systems, and workers need a management layer. The question everyone should take back to your laptop: in my own agent stack, what happens on the second concurrent write, and can I prove, today, under whose authority my agent acted?

Special Thanks

We will be giving wrist bands to attendees.
When coming, please make sure to have your Luma account or your Meetup account ready on your laptop web browser, or on the respective mobile app.

Tags: Technology, AI Agents, Tool-Calling, Database

Update: In the later part of July, we will send out a question for you to indicate your food preferences (chicken, or beef, or not meat) which will affect the food ordering for the dinner.
So please keep a lookout for it and respond to it when it is sent :)

Hosts

AINight
SQ Collective

Topics

Panel / Roundtable