I see data as the particles of our digital universe

Shaping the world around us as we gather, connect, build, and grow with it

Building products that deliver real impact

About

What I do, and how I decide what's worth building

I'm a Product Manager based in London. Six years across AI, SaaS and e-commerce: three years owning an ML recommendation product used in production by 1,000+ merchants, then six months taking a B2B2C SaaS platform from problem definition through prototype and MVP to a live product, as its only PM. Most recently, I led growth and AI automation for a UK platform with 127k+ monthly active users.

What I do is find the constraint. At Rosetta.ai, the question was never whether the model was accurate. It was which client segment would actually benefit enough from personalisation to make it worth rolling out, before we spent a year taking it to everyone. At Banyan and JUKSY, the ceiling on growth wasn't traffic; it was how much content one editor could produce in a day.

I focus on the result, working hands-on with data, model evaluation and agentic tooling to test product directions before engineering commits to them. I also build and run my own products on the side to keep that honest. Most AI products fail not in the modelling, but in translating it into a commercial outcome. That translation is the part I do.

  • 01Clarify goals and constraints first

    Every roadmap I've owned started by naming the one thing actually capping the business, whether that was throughput, margin or activation, rather than working down the longest list of feature requests. Define the problem, break it into parts, then decide what gets built and in what order.

  • 02Ship the smallest version, then scale it

    I'd rather put a working prototype in front of real users than defend a hypothesis in a document. At JUKSY that meant a prototype, then an MVP live in the market, then feature work shaped by user feedback and test results rather than by what we assumed they wanted, with a metric agreed before launch and an honest read on it after.

  • 03Automate the repetitive work

    Any process that runs the same way twice stops being product work and becomes something to automate. Doing that at Banyan doubled output per editor at flat headcount and gave the team its time back for the parts that needed judgement.

Selected Work

Three case studies across AI, SaaS and e-commerce | Tap any project to open the full case

01AI AutomationGrowth2026

Banyan Digital

AI Automation & Growth

Ran product and operations for a social platform with 127k+ monthly active users. Diagnosed content production as the real constraint and shipped five AI automation workflows that doubled output per editor-day at flat headcount.

127k+
Monthly active users
5 → 10
Pieces per editor-day
5
AI workflows shipped
9m 30s
Avg engagement time
02B2B2C SaaSSole PM2024-2025

RoBard SaaS

Marketing Automation Platform

As the sole product manager, I built this marketing SaaS platform from zero to one, covering AI-assisted asset collection, AI copywriting trained on brand voice, one-click multi-channel publishing and KOC/KOL marketing. I led a cross-functional team of 10+, shipping an MVP in three months and a live product with a fast iteration cycle within six.

10+
Cross-functional team
3 months
MVP shipped
7 → 20
Articles per editor-day
+80%
Ops efficiency

Team of 10+ across product, engineering, design, editorial, sales and finance.

03AI RecommendationE-commerce2021-2024

Rosetta.ai

AI Recommendation Engine

Built an AI personalisation engine with the machine learning team and deployed it across 1,000+ e-commerce clients, turning generic bestseller widgets into real-time, per-user recommendations.

1,000+
E-commerce clients
+660%
Revenue increase
+350%
CVR improvement
-43%
Inference latency

Commercial figures are first-year averages for early-adopter accounts migrating from zero personalisation. Latency is measured against the previous production solution.

Side Projects

Products I scoped, built and shipped on my own

04Live productAgentic automation2026

Credit-card rewards search

Card rewards shift constantly and spending contexts keep multiplying, so almost nobody tracks which promotions are live at any given moment. The result is the same question at every checkout: which card do I use to get the most back? I mapped the common spending contexts against 30+ cards, so you register the cards you actually hold once and get the best-return answer on every purchase after that.

30+
Cards in the index
Daily
Autonomous data refresh
Solo
Product, build, ops
Where are you spending?Live
NetflixFlightsShopeeCostcoApple
+8.5%E.SUN U Bear
+4.5%SinoPac Dual Currency
+3.0%Cathay CUBE
+2.0%Fubon Costco Card
-0.2%Taishin Richart
Net reward after FX fees, caps & registration30+ cards indexed
Live site ↗ Open the tool ↗
05LLM productUnit economics2026

Daily Brief

Personalised morning briefing bot

A Telegram bot that delivers the day's news at a set time, filtered by the industries you care about and capped at however many stories you want. It also handles learning tracks: pick a subject like accounting or finance and it builds a daily study plan alongside the briefing.

The constraint
The interesting part was the cost. Running an agent through several WebSearch calls and having it read each result end to end burns tens of thousands of tokens on search output alone.

Solution
Switching to a script that pulls RSS first, returning forty headlines and summaries as cheap plain text, then handing the model a single select-and-write pass, cut token use by five to ten times and made the output more consistent.

Daily briefDelivered
07:30 your chosen time
Scanning sources…
Reading 40+ feeds, then picking yours
World
Markets open higher after the policy shift
London jobs
12 new product roles matched your profile
Finance literacy
What a base-rate cut does to your savings
Because you follow
AI tooling | outdoor gear | K-pop
Written per readerFlat cost curve
Open the bot ↗ Open the bot ↗
06Data productScoping under constraint2026

Travel Radar

Flight price agent

Built for getting from A to B at the lowest real cost, covering both direct routes and self-connected multi-leg ones, across fixed or open date ranges. Ask for ten days in Iceland leaving from Paris any time in January, and Travel Radar prices the direct flights alongside the alternatives: Paris to London to Reykjavik, or any other city worth routing through. Not just fixed dates, and not just what the airlines sell as a single connecting ticket.

Comparing faresWatching
LHRAMSTPE
LGWISTBKK
STNHELICN
✓1 stop | split ticket£468
Direct£742
✓Direct£512
1 stop | split ticket£559
✓1 stop | 2 nights in HEL£431
Direct£688
Alerts only below your thresholdAny origin | any destination
Open the bot ↗ Open the bot ↗

Professional Experience

  1. Apr 2026 - Jul 2026

    Product Operations Specialist

    Banyan Digital Limited | London, UK | Hybrid

    Ran a content social community platform with over 127k monthly active users. Through data analysis and stakeholder interviews across departments, I designed and shipped five automated workflows that doubled output from 5 to 10 pieces per editor-day.

  2. Apr 2024 - Aug 2025

    Product Manager

    Juksy Digital Media Co., Ltd. | Taipei, Taiwan | Hybrid
    Promoted from Junior Product Manager within nine months

    Built a marketing automation SaaS and an AI content pipeline that lifted editorial output from 7 to 20+ articles per editor per day. Ran agile delivery across a cross-functional team of 10 spanning engineering, design and sales, reporting to C-suite stakeholders.

  3. Feb 2021 - Apr 2024

    Product Specialist & Domain Expert

    Rosetta.ai | Taipei, Taiwan | Hybrid

    Built AI recommendation and personalisation for fashion e-commerce with the ML team, deployed across 1,000+ merchants. Owned and built the 10,000+ attribute taxonomy and database that the ranking model scored against. Worked with engineering and data science to cut response time by 43%.

  4. Oct 2020 - Mar 2023

    Project Manager

    Freelance | Taipei, Taiwan

    Coordinated projects across industries, managing client and supplier relationships alongside e-commerce platform operations. Delivered two films, three music videos, one documentary and one curated event, plus a range of ongoing business partnerships.

  5. Mar 2020 - Aug 2020

    Procurement Specialist

    MAKALOT Industrial | Taipei, Taiwan | On-site

    Managed supplier orders across China, Hong Kong and Southeast Asia, covering pricing negotiation, quality control and delivery scheduling. Ensured the on-time launch of 8 apparel products for Japanese fast-fashion brand GU.

  6. Sep 2015 - Jun 2019

    BA Fashion Marketing

    Fu Jen Catholic University | Taiwan

    Consumer and market analysis, project management and go-to-market strategy. Winner, 46th Graduation Exhibition; led the end-to-end marketing campaign for the graduate degree show.

Capabilities

Product judgement, AI fluency and the technical depth to ship with engineering

AI & Data

  • LLM integration
  • Agentic workflow automation
  • Model evaluation and guardrails
  • A/B testing and experimentation
  • SQL
  • Google Analytics
  • SEO
  • Claude Code
  • Codex

Product

  • Product strategy
  • Roadmapping
  • Discovery and user research
  • PRD and functional specs
  • Prioritisation
  • Prototyping and UX
  • GTM and launch
  • Agile and sprint planning

Tools

  • Figma
  • Jira
  • Miro
  • Notion
  • Slack
  • LINE OA

Leadership

  • Cross-functional delivery
  • Stakeholder management
  • C-suite communication

Domains

  • SaaS
  • E-commerce
  • Fashion business
  • Digital media

Languages

  • Mandarin and Taiwanese (native)
  • English (fluent)

Let’s build something great together