Andrew Chew
Singapore · San Francisco

From mandate to machine.

AI-native Strategy & Operations.

I take an ambiguous mandate, close the deals and get the sign-offs it needs, then build the process, the team or the agent that runs it without me. Ten years in strategy and operations across consulting, big tech, crypto and fintech.

See the work LinkedIn
Mandate
Machine
Ten years of strategy & operations at
BCG Uber Meta Gemini Airwallex
Studied and worked in the US and Singapore · San Francisco · New York · Singapore

Select commercial achievements

Deals closed, supply run, operations owned. Each one needed the same three things: understand the workflow, find the edge cases, get every function to yes.

What I do

Three things, usually at the same time.

Commercial

Partnerships and deal structuring

Value exchange, product alignment on both sides, contract redlining, signature, and the operating metrics that prove the deal worked. I have done this between two large product organizations and know where deals die.

Operations

Operations strategy in regulated fintech

Payouts, deposits, card issuing, partner banks, regulators. I know who has to sign off inside a fintech before anything ships, and how to get them there.

Build

AI-native operations

I analyze the workflow first, then build the agent or automation that runs it. Claude Code, Codex and Cursor day to day; an AI support agent, a property-management stack and a regulatory pipeline in production.

Case studies

Three systems I built or shipped. Same format each time: the situation, what I did, what happened.

Leading fintech companyAI agentCustomer supportIn production

Deployed an AI support agent that resolves payout, deposit and global-account tickets on its own.

Situation

Support ran on note-based Zendesk workflows. Ticket volume for payouts, deposits and global accounts kept rising and nobody could say why, because nothing analyzed the tickets.

The hard part was not the tooling. A payout failure over email cannot be explained the way it can in chat, because email can't verify who is asking.

What I did

  • Used Claude Code and Codex to analyze the Zendesk ticket base first, so we automated the right things.
  • Built the triage taxonomy across payouts, deposits and global accounts.
  • Worked directly on an AI support platform to move the agent from node-based flows to natural-language workflows.
  • Designed separate handling per channel, including the rule that email payout-delay inquiries never disclose failure details.
  • Iterated on three numbers: tickets covered, solve rate, CSAT.

What happened

The agent resolves tickets independently across internal chat and email, with CSAT held or improved.

Coverage: 57% of payout, deposit and global-account tickets. Solve rate: 68%. CSAT: 63% → 84%.

The taxonomy and channel rules are now the spec other teams build against.

Role: owner, build and rollout·Stack: AI support platform, Zendesk, Claude Code, Codex
Leading fintech companyRegulatory reportingClaude CodeCard issuing

Turned card-issuing reports for a Netherlands financial regulator into code, edge cases included.

Situation

Monthly and quarterly reports to a Netherlands financial regulator, among the strictest in Europe, covering payouts, deposits and card issuing across card and non-card transactions. Each cycle was rebuilt by hand from the regulator's documents.

The edge cases are the whole job: what counts as a fraudulent versus a rejected transaction, which stage of a card transaction gets counted, whether Apple Pay and Google Pay are card transactions, and how Visa and Mastercard fraud reason codes map to the regulator's categories.

What I did

  • Ran three workstreams: business logic, data validation, and reporting automation.
  • Processed the regulator's documentation and wrote the counting rules down as rules, one per edge case, before touching data.
  • Validated the data logic against the issuing pipeline: transaction counts and volumes in the Netherlands, transaction stages, wallet tokens.
  • Built the report generation with Claude Code so every number is reproducible from the same rules.

What happened

The report runs from documented rules instead of memory. Every edge case has a written decision that finance, compliance and engineering can see.

Cycle time: 17 days → 14 hours. Corrections after filing: 23 → 11.

Role: sole builder·Stack: Claude Code, SQL, the regulator's own documentation
PersonalUS real estateOperationsClaude integrations

Run a US rental real estate business from Singapore on two to three hours a week.

Situation

Two properties in Texas and Indiana, bought and renovated while I was working in San Francisco, listed on Airbnb and Vrbo.

In 2025 I moved back to Singapore. Fourteen hours ahead, a full-time job, and guests who message at 2am.

What I did

  • Wrote the operating process end to end, then hired a Philippines-based assistant to run it.
  • Property-management system automating guest messaging across both platforms.
  • Dynamic pricing tool for rates; a business phone line for guests and vendors.
  • Connected the pricing tool and the phone line to Claude so pricing, messages and issues get reviewed in one place instead of four apps.

What happened

The business runs on two to three hours of my time a week. The assistant handles the day to day; I handle pricing decisions and exceptions.

Occupancy: 73%. Revenue: cash flow positive.

The smallest system I run and the clearest example of how I think: process first, people second, tools third.

Role: owner and operator·Stack: Guesty, PriceLabs, Quo, Claude

Journey

Singapore, two years of college in the US, two years across Southeast Asia in consulting, then five years in San Francisco, then back home.

Let's talk.

Building something where the playbook isn't written yet? I'd like to hear about it.

Message me on LinkedIn