Web studio, co-founded

TechStarterWeb: a studio website that is also its order desk

TechStarterWeb Consulting builds personal and small-business websites for clients in Taiwan, Hong Kong and Singapore. I co-founded it in October 2026 and lead the technical side. The studio's own site, live since 23 October 2026, is its first piece of work.

Astro · Cloudflare Workers · Cloudflare D1 · Cloudflare Access · OpenCC

The need

One site for clients and for the studio

The public pages take enquiries in three languages; a back office behind sign-in turns each one into an order with a status and a log of every change.

How an enquiry moves
  1. Enquiry
  2. Stored
  3. Sign-in
  4. Status
  5. Logged
  1. Enquiry
  2. Quoted
  3. Signed
  4. In production
  5. Delivered

In use

Private by design, and live on its own domain

Only the team's addresses can sign in to the back office, through Cloudflare Access with one-time email codes. The site is live on its own domain, with Google Search Console and visit statistics.

Limits

What it does not do yet

Enquiries are not yet emailed to the team, and Cloudflare's free plan has no servers in mainland China, so some networks load the site slowly or not at all.

How it was built

Built with an AI coding agent, in five layers

I set the requirements and review each change; the agent writes the code. Routine work goes to cheaper models: ten calls to the low-cost model came to about US$0.03.

How the work is organised
  1. L0 People decide, pay, merge
  2. L1 Lead agent splits work into task cards, reviews
  3. L2 Worker agents write code, open pull requests
  4. L3 Low-cost model copy, translations, summaries
  5. L4 Scripts and CI anything a rule can decide

Use the lowest layer that can do the job

How it stays on track

Task cards and checks keep every change in bounds

Each task card lists its goal and the files it may change, and CI fails a pull request that breaks the rules.

Full write-up on GitHub

The life of a task card
  1. To-do
  2. Task card
  3. agent:ready
  4. Agent run
  5. Pull request
  6. CI checksfile locks · version numbers · spec sync
  7. Merge
  8. Auto wrap-up
    Business Analytics coursework

    Mobile money adoption in the Global South

    An individual Tableau story, submitted in December 2025: five dashboards on how mobile money is spreading across the Global South, built from World Bank Global Findex and GSMA data.

    Tableau · Excel · Python · World Bank Findex and GSMA data

    The data

    From 437 coded fields to 24 measures

    An extract of 437 coded Findex fields across 174 economies was mapped down to 24 business-readable measures. The 2011 and 2022 waves were dropped for thin coverage, leaving 171 economies.

    Data preparation
    • Coded fields to measures 43724
    • Economies 174171
    • Survey waves
      1. 2011 (dropped)
      2. 2014
      3. 2017
      4. 2021
      5. 2022 (dropped)
      6. 2024

    The story

    Five dashboards, each with an insight and a recommendation

    A cover opens the story, five interactive dashboards follow, and a conclusion closes it.

    The story, in order
    1. Cover
    2. Mobile Money Boom
    3. Leapfrogging the Plastic Era
    4. The Catch-Up
    5. The Usage Paradox
    6. Platform Transformation
    7. Conclusion

    What it found

    Mobile money leads in Sub-Saharan Africa

    Adoption expanded steadily from 2014. Digital payment use runs ahead of debit-card ownership, and savings trail both.

    Limits

    What the data cannot show

    My critical assessment of each dashboard names the limits: no transaction-level detail and no qualitative evidence.

    Where it led

    It led to my final-year research

    Digital payment use running ahead of debit-card ownership raised the question my final-year project tested: what happens to banks' fee income as payments move away from cards.

    Back to the final-year research
      Open the interactive story on this page
      Scaled to fit this screen. The Tableau Public story link above opens it at full size.
      Open source

      layered-ai-dev-workflow: how the studio works with AI coding agents

      The layered way of working that TechStarterWeb uses with AI coding agents. The five-layer figure is in the TechStarterWeb project above.

      The idea

      A lead agent plans, worker agents build, CI checks

      A lead agent turns a to-do into a task card, worker agents write the code, and scripts and CI check whatever a rule can decide.

      From a to-do to a checked pull request
      1. To-do
      2. Task card
      3. Pull request
      4. CI checks

      State

      State lives in GitHub, not in chat

      Work state is kept in issues and handoff notes, so a new session can pick up where the last one stopped.

      Rules

      Rules are enforced by checks

      CI fails a pull request that edits files outside its task card, reuses a version number or skips the spec update.

      CI checks on every pull request
      • File locks
      • Version numbers
      • Spec sync

      Reuse

      Open source, with tests

      Published with tests, so that others can reuse or adapt it.

      Full write-up on GitHub

        Graduate research

        Digital payment pressure and bank net fee margins

        A panel of 39 commercial banks in Indonesia, Malaysia and Thailand, 2014–2022. The log specification shows a negative, statistically significant pooled association that comes mainly from differences between countries: a conditional association, not a causal effect.