✕

Product Work

AI product work in high-stakes environments.

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 ↓ About & Contact →

Featured Work

Selected case studies across AI, Trust & Safety, and Data Platform

Escalation by Design: Multi Agent Fact Checking with Policy Constraints

Confidence thresholds plus human routing reduce false positives at scale.

Read case study →

Discovery Frameworks and ROI Modeling for Procurement AI

Buyer shadowing and RAG design that reduced search time by 25%.

Read case study →

Four Layers of Harm Mitigation in Enterprise AI

Model choice, grounding, and UX constraints for safer enterprise AI.

Read case study →

AI Architecture

RAG in Practice: Where It Works, Where It Fails, and What to Use Instead

A product-first guide to where RAG fits and where other patterns win.

Read case study →

All Product Work

From One API Call to a Production Agent

Sequencing Architecture Layers in Enterprise AI Copilot Systems
ENTERPRISE
Posted on March 21, 2026

I spent months building enterprise AI copilots in a large automotive setting (including Volvo): systems that draft customer-facing content, triage incoming documents, and assist compliance workflows across markets. Early on, I read Anthropic’s Effective harnesses for long-running agents and Effective context engineering for AI agents, treated them as build instructions,... [Read More]
Tags:
Building a Self Improving AI Assistant from 10 Markdown Files

Local First Assistant Design with Auditable Evolution and Controlled Complexity
0→1 BUILDS
Posted on February 27, 2026

OpenClaw hit 230,000 GitHub stars in early 2026 by treating AI assistants as an infrastructure problem: persistent identity, structured memory, tool orchestration, and 24/7 uptime across 50+ messaging channels. The architecture is impressive. It is also far more than I need. [Read More]
Tags:
RAG in Practice: Where It Works, Where It Fails, and What to Use Instead

System Boundaries, Failure Modes, and Architecture Choices for RAG
ENTERPRISE
Posted on February 6, 2026

I have built RAG systems for enterprise knowledge retrieval in regulated domains: compliance documentation, trade policy, and product specifications across multiple markets. The recurring pattern is simple. Teams treat RAG as solved once the demo works, then hit failure modes in production where error cost is highest. [Read More]
Tags:
Escalation by Design: Multi Agent Fact Checking with Policy Constraints

Confidence Gating, Policy Separation, and Human Routing in Content Moderation
ENTERPRISE
Posted on January 27, 2026

I built a proof-of-concept for policy-aware content moderation: a multi-agent pipeline where AI proposes, policy constrains, and humans decide. The demo is live at llm-misinformation.streamlit.app. [Read More]
Tags:
Breaking Changes on a Deadline in Enterprise EDIFACT Integration

Regulatory Cutover Execution Across Three Countries with Production Reliability
ENTERPRISE
Posted on April 24, 2025

In Germany’s deregulated energy market, utilities, grid operators, and metering companies exchange millions of structured messages every month: meter readings, supplier switches, billing data, and master data updates. The messaging format is EDIFACT, governed locally by regulatory and industry bodies such as BNetzA and BDEW. When these bodies publish a... [Read More]
Tags:
The Semantic Layer in Modern Data Architecture

Why Modern Data Stacks Need a Governed Semantic Layer
ENTERPRISE
Posted on February 13, 2025

At Volvo Cars, “duty savings” had 23 conflicting definitions across 50+ teams. Finance calculated it one way for the earnings call. Operations calculated it differently for performance reviews. Both were locally correct. Both produced different numbers in the same executive meeting. [Read More]
Tags:
Discovery Frameworks and ROI Modeling for Procurement AI

Buyer Shadowing, Cognitive Load Measurement, and a 25% Search Time Reduction
ENTERPRISE
Posted on February 2, 2025

Procurement at Volvo Cars manages 1,169 suppliers, over 7,000 contracts, and millions in annual spend across three fragmented legacy systems: VGS, VPC, and SI+. When I joined this initiative, there was no product brief. The executive mandate was AI exploration for procurement. The growing backlog of buyer complaints pointed to... [Read More]
Tags:
Photo to Listing in 7 Seconds: Unit Economics and Emotional Design in Mobile AI

Unit Economics, Activation Friction, and Emotional Design in a Solo Built Resale App
0→1 BUILDS
Posted on January 24, 2025

SnapSell turns a single photo into a ready-to-post resale listing in under 10 seconds. I built it solo to solve a problem I lived with: spending 5 to 10 minutes writing each listing across Depop, Facebook Marketplace, and Poshmark, and often abandoning the process because the text-writing step felt disproportionate... [Read More]
Tags:
Four Layers of Harm Mitigation in Enterprise AI

Model Selection, Grounding, and UX Constraints for Trustworthy Procurement AI
ENTERPRISE
Posted on January 17, 2025

Over the past year building generative AI inside Volvo Cars’ procurement organization, I confronted a design problem that most AI product literature skips: how do you make an inherently uncertain technology behave predictably in an environment where mistakes carry legal, financial, and compliance consequences? [Read More]
Tags:
Moving Failure Upstream for SWIFT Validation at Airwallex

Corridor Aware Validation and Precision vs. Conversion Control in SWIFT Payments
ENTERPRISE
Posted on January 15, 2025

Airwallex invests heavily in local payment rails (ACH, SEPA, Faster Payments) to make global payments feel local. For most corridors, this works. For long-tail destinations, specific currencies, and regulated markets, SWIFT is the only viable option. [Read More]
Tags:
Turning Federated Data Chaos into Controllable Decision Systems

Risk Framing, Decision Boundaries, and Federated Governance at Scale
ENTERPRISE
Posted on June 11, 2024

At 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]
Tags:
Circadian Science for Personalized Sleep Plans

Turning Temperature Minimum (T-Min) Research into Age Constrained Sleep Planning
0→1 BUILDS
Posted on February 1, 2024

When my family faced a long-haul flight with a 3-month-old, I realized 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 once.... [Read More]
Tags:
akaTask: From Ambiguous Parenting Pain Points to a Shipped AI Task Manager

Structured Framing, Tradeoff Discipline, and Scope Control for Product Launch
0→1 BUILDS
Posted on January 5, 2024

I turned a messy, emotionally charged problem space into a shipped iOS app. The research that informed it is covered in Navigating Parenthood: Product Preferences and Parental Challenges. Starting where research ended: translating validated insights into a product with clear scope, measurable targets, and deliberate constraints on where AI belongs... [Read More]
Tags:
Mapping Cognitive Load to AI Feature Bets from Parent Interviews

Interview and Survey Evidence Turned into a Scoped Parenting AI Product
0→1 BUILDS
Posted on January 3, 2024

The starting point was a sentiment, not a product: new parents are overwhelmed, and AI could help. That framing is useless for product work. “Overwhelmed” spans hundreds of micro-problems across feeding, sleep, purchases, social isolation, and daily logistics. AI is a capability, not a solution. The real question was which... [Read More]
Tags: