About
Financial decisions stop being automatic where the evidence moves into data you cannot reach or documents you cannot trust. My work has been moving that point, safely.
I keep coming back to three questions. Access: can we get the information? Understanding: can we turn it into facts? Trust: will an institution act on those facts? Better models change the answers. They also change which question becomes the constraint.
Explore with AI
Ask an AI about my work, or use this site’s read-only tools through the Model Context Protocol (MCP).
A model can answer one question while another still holds up the work.
- Access Can we reach the evidence?
A missing source leaves the decision incomplete.
Find the source before using the answer.Dragon Wealth: relevant fund data was difficult to obtain.- Understanding Can the source support this fact?
A material qualifier can remain unresolved after extraction.
Expose the uncertainty and route the judgment.Document AI: the field that changes an obligation needs its own check.- Trust Should this decision be delegated?
Available evidence does not settle permission or the consequences of error.
Set the scope, owner and checks for the decision.Finaxar: larger credit lines went to manual review.
I. Bank risk: what an institution can audit
At Deloitte and Numind, I worked on bank risk and strategy across Europe, Asia and the Middle East. The starting point was the evidence a risk team or regulator needed before trusting a decision. That remains my first principle: institutions trust what they can audit.
II. Crayon Data and Dragon Wealth: the facts underneath the model
At Crayon Data, the model was a small part of a billion-node graph. At Dragon Wealth, fund documents and complex structured data became personalised recommendations and model portfolios, delivered through web and mobile applications. A wrong fact at the start could corrupt everything built on it.
The access problem persisted even when the model worked. Retail-relevant fund data was difficult to find; portfolios above roughly US$200,000 still needed manual optimisation, especially for tax. Ground truth comes first.
III. Finaxar: the stakes move the boundary
At Finaxar, we made instant credit decisions through partnerships with Lazada and Stripe. Larger SME lines needed evidence the platforms could not supply through APIs. Bank statements and PDFs brought the work back to extraction and human review. Lines above roughly US$20,000–25,000 went manual.
I led a lending and payments business across four markets: Singapore, Hong Kong, Vietnam and Malaysia, with a team that reached 100+ people across five countries. The decision to automate had to account for both the evidence available and the money at risk. Autonomy is earned per decision, and depends on the stakes.
IV. Menyala and Splore: relevance needs a source
From July to October 2022, during Splore’s incubation at Menyala, we tested a live semantic-search MVP with real users. That preceded Splore’s formal kickoff in November. The work moved from consumer information to unstructured content and enterprise financial services, then narrowed to private markets.
The access problem I had encountered at Dragon Wealth returned at institutional scale. Retrieval could find something relevant; an answer also needed evidence a reader could inspect. Splore developed source-cited answers and was acquired by Model ML. Every answer points to its source.
V. Regulated document AI: know when to hand back the decision
In regulated document work, an average accuracy score can hide an error in the field that determines a financial obligation. The questions become more precise: what is the ground truth for this field, when should the system abstain, and which regression should stop a release?
The principle is abstain and route; gate releases on critical fields, not averages. Human review is part of the operating design, with attention directed to the decisions that need it.
VI. Agent Office: responsibility for actions
My independent Agent Office explores the same problem when agents take actions. A continuing obligation needs an owner after the conversation ends. Permission to use a tool and evidence that its action worked answer different questions.
Autonomy is earned per action. I want to understand when delegation returns attention, including how often I have to take the work back. The office is a personal system for projects and independent research.
Where it leads
I use earned autonomy as a design principle: the scope of an agent's work should grow with evidence that it can handle the task reliably. Agent Office is where I explore that principle in practice.
The question ahead is how far agents can extend reliable automation in regulated operations, and what evidence would justify giving them the next decision.
Read each cell as a proposed choice, with evidence assessed for the specific task.
- Easier to reverse · Incomplete evidence
- Gather or draft
Keep the action reversible while resolving missing evidence.
- Easier to reverse · Sufficient evidence for the task
- Delegate within scope
Define the expected result and check it.
- Harder to reverse · Incomplete evidence
- Pause the action
Resolve the evidence and name the decision owner.
- Harder to reverse · Sufficient evidence for the task
- Review before acting
Keep approval and recovery responsibility explicit.
Task-specific permissions and the cost of checking apply in every cell.
Career
- Based in Singapore; I lead the Data & AI business unit at Marketnode as Vice President, Head of Data & AI, since 1 November 2025.
- Previously: Splore, CEO & Director, February 2023–2025 (project kickoff November 2022); Menyala, Resident Builder, July 2022–January 2023.
- Finaxar: Co-Founder, May 2016–July 2022; CEO to January 2021, then Chairman.
- Dragon Wealth Asia: CTO & Head of Product, May 2014–April 2016; Operating Advisor at Monk’s Hill Ventures, October 2014–October 2022.
- Avagam: Founder & CEO, July 2010–March 2012; Crayon Data: Technology Leader, April 2012–April 2014.
- Earlier: Deloitte, risk management and banking consulting, 2005–2009; Numind, strategy consulting, 2009–2010.
- Education: Master of Science and Bachelor of Science (Honours), Mathematics & Computing, IIT Kharagpur. INSEAD Global Management Programme (CGM), alumnus; Harvard Business School High Potentials Leadership Program (executive education).