Hi, I'm Cynthia.
I'm a Product Manager who builds AI decision systems for high-stakes, policy-constrained environments, combining LLM-powered workflows with human-in-the-loop governance and measurable quality signals.
I've shipped products at Volvo Cars, SAP, Airwallex, and Alibaba, spanning procurement AI, trade compliance, cross-border payments, and data platforms across Europe, Asia, and the U.S.
Featured Work
Selected case studies across AI, Trust & Safety, and Data Platform
Escalation by Design: Multi-Agent Fact-Checking
Confidence gating & human routing to reduce false positives in content moderation at scale.
Read case study →Procurement AI: Discovery & ROI Modeling
Shadowing buyers to quantify cognitive load, then building RAG that cut search time 25%.
Read case study →Four-Layer Harm Mitigation in Enterprise AI
How model selection, grounding, and UX constraints combine to build trustworthy AI systems.
Read case study →RAG vs Context Engineering: Insights & Limitations
Why retrieval-augmented generation is not universal: a product-first decision framework.
Read case study →All Product Work
-
In early 2026, an open-source project called OpenClaw exploded onto the scene, surpassing 230,000 GitHub stars within weeks. It represented a fundamental shift in how we think about AI assistants: not as stateless chatbots, but as persistent, context-aware digital coworkers that run on your own hardware. [Read More]
-
RAG vs Context Engineering: Insights and Limitations
Why retrieval-augmented generation is not universal
ENTERPRISERetrieval-Augmented Generation (RAG) has become the default architecture for grounding LLMs in enterprise knowledge. The industry conversation, however, often conflates “using RAG” with “solving the knowledge problem.” These are not the same thing, and treating them interchangeably leads to brittle systems, wasted investment, and false confidence in production. [Read More] -
Escalation by Design: Multi-Agent Fact-Checking with Policy Constraints
How confidence thresholds and human routing reduce false positives in content moderation
ENTERPRISEContent moderation at scale is one of the hardest problems in modern tech. When platforms like TikTok, YouTube, or Facebook need to make decisions about billions of pieces of content daily, they cannot just build a better AI model and call it solved. The real challenge is not only detecting... [Read More] -
The Semantic Layer in Modern Data Architecture
Why data lakes, lakehouses, and meshes still fail without a governed semantic layer
ENTERPRISEAt Volvo Cars, data is generated at enormous scale across manufacturing lines, supply chain systems, dealer networks, connected vehicles, and procurement platforms. The challenge is not producing data. The challenge is producing meaning. Across 50+ teams, the same metric was calculated differently. “Duty savings” alone had 23 conflicting definitions. Finance... [Read More] -
Discovery Frameworks and ROI Modeling for Procurement AI
Shadowing buyers to quantify cognitive load, then building RAG that reduces search time by 25%
ENTERPRISEProcurement at Volvo Cars manages 1,169 suppliers, over 7,000 contracts, and millions in annual spend distributed across three fragmented legacy systems: VGS, VPC, and SI+. When I joined this initiative, there was no product brief. What I had was a vague executive mandate to “explore AI for procurement” and a... [Read More] -
Photo-to-Listing in 7 Seconds: Unit Economics and Emotional Design in Mobile AI
Balancing inference costs, activation friction, and delight in a solo-built resale app
0→1 BUILDSSnapSell is a solo-built case study in managing AI quality risk within constrained economics. The product transforms a single photo into a complete resale listing, but the core challenge was designing a system that balances inference costs, hallucination risk, and user trust without human review infrastructure. I designed every decision... [Read More] -
Four-Layer Harm Mitigation in Enterprise AI
How model selection, grounding, and UX constraints combine to build trustworthy procurement systems
ENTERPRISEOver the past year building generative AI capabilities inside Volvo Cars’ procurement organization, I confronted a fundamental challenge: how do you deploy generative AI in a high-stakes environment where mistakes carry legal, financial, and compliance consequences? The answer wasn’t better models or more features. It was turning an inherently uncertain... [Read More] -
Moving Failure Upstream for SWIFT Validation at Airwallex
Corridor-aware validation that balances precision vs conversion to reduce support load by 25%
ENTERPRISEAirwallex’s core promise is to make global payments feel local. The company invests heavily in local payment rails (ACH, SEPA, Faster Payments), multi-currency accounts, and a unified internal ledger powering collections, FX, payouts, and cards. From the outside, this looks like a clean, modern alternative to traditional banks. [Read More] -
Turning Federated Data Chaos into Controllable Decision Systems
How I framed risk, designed decision boundaries, and scaled governance without breaking autonomy
ENTERPRISEAt Volvo Cars, dozens of domain teams gained the ability to publish data products independently under a federated data mesh architecture. Development velocity increased dramatically. So did the rate of silent failures. [Read More] -
Circadian Science for Personalized Sleep Plans
Translating temperature minimum (t-min) research into personalized, age-constrained sleep plans
0→1 BUILDSWhen my family faced a long-haul flight with a 3-month-old, I realized that every piece of jet lag advice I could find was designed for solo business travelers. Nothing addressed the cascading failure that happens when a baby, a toddler, and two exhausted parents all lose their sleep anchors at... [Read More] -
akaTask: From Ambiguous Parenting Pain Points to a Shipped AI Task Manager
How structured problem framing, deliberate tradeoffs, and disciplined scoping turned messy user research into a live product
0→1 BUILDSNew parents face a sprawling, emotionally charged problem space: sleep deprivation, information overload, shifting routines, and constant decision fatigue. When I started exploring this space, there was no single well-defined problem to solve, only a tangle of anxieties, fragmented workarounds, and strong opinions from every direction. This case study walks... [Read More] -
Mapping Cognitive Load to AI Feature Bets from Parent Interviews
How 20 interviews and 150 surveys turned ambiguous parenting pain into a scoped AI product with measurable time savings
0→1 BUILDSThis project began with a vague hypothesis: new parents are overwhelmed, and “something with AI” could help. That is not a product. It is a sentiment wrapped in a buzzword. The real work was figuring out which problems were acute enough to warrant a new tool, where AI would genuinely... [Read More]