Harvey Singh, AI engineer

I ship production systems, not prototypes.我交付的,是真正上線運作的系統,而非原型。

Six years across FinTech, SaaS and EdTech. Now building LLM applications, agent tooling and automation.六年橫跨金融科技、SaaS 及教育科技。現正專注於 LLM 應用、Agent 工具與自動化。

Portrait of Harvey Singh
professionals trained
0+
projects shipped
0+
users on The Asset app
0+
shipping production software
0 yrs
  • Next.jsNext.js
  • ReactReact
  • TypeScriptTypeScript
  • PythonPython
  • LaravelLaravel
  • SwiftSwift
  • KotlinKotlin
  • FlutterFlutter
  • Tailwind CSSTailwind CSS
  • PostgreSQLPostgreSQL
  • MySQLMySQL
  • SupabaseSupabase
  • FirebaseFirebase
  • Node.jsNode.js
  • VercelVercel
  • DockerDocker
  • three.jsthree.js
  • ExpoExpo
  • AnthropicAnthropic
  • Hugging FaceHugging Face
  • TensorFlowTensorFlow
  • StripeStripe
  • n8nn8n
  • GitHubGitHub
  • PrismaPrisma
  • SentrySentry

Selected work

Production systems with real users, and the decisions behind them. Each one opens into a short case study.

Experience

  1. AI & Software Engineering Instructor at Preface Nomad

    Jul 2020 to present

    • Trained 380+ professionals in applied AI and software engineering, from LLM tooling and machine learning to full-stack and mobile delivery.
    • Took 114 trainee projects from concept to live deployment across AI/ML, web, games, mobile, generative media and data science.
    • Supervised on-device Llama inference, a CNN trained from scratch for live gesture recognition, and churn prediction models.
    PythonNext.jsTensorFlowLlamaUnitySwiftUI
  2. Lead Full-Stack Developer at The Asset

    Jun 2022 to Jun 2025

    • Originated The Asset's mobile app and led it from first idea to both app stores, growing from sole developer to leading a team of four. Past 5,000 users by September 2024, an audience of bankers, investors and corporate treasurers rather than consumers.
    • Replaced a paper nomination process with a live voting platform where buy side and sell side rate each other, feeding the Asset Benchmark Research data that adjudicates the Triple A awards. Designed the integrity rules myself: tiering, weekly vote caps and the points model.
    • Designed the MySQL schemas and the business logic behind voting, awards, events and content, and the UI and UX in Figma.
    • Digitised the magazine, podcasts and video into the app, with a feed ranked against each reader's stated professional interests.
    • Built the event platform end to end: registration, QR check-in, live polling, Q&A and in-room games, plus newsletter runs to 100,000+ subscribers.
    • Owned the commercial and operational surfaces too: sponsor and ad placement on the Laravel site, Firebase and Google Analytics instrumentation, vendor management, and RPA that removed recurring editorial work.
    LaravelLivewireMySQLAWSSwiftUIJetpack ComposeFirebaseFigma
  3. Full-Stack Developer, Mobile & Web at TheCodingMachine

    Sep 2021 to Mar 2022

    • Built Geowitness, a location-first video app now live on the App Store, in Flutter and Dart against a Symfony and GraphQL backend, with a Vue admin panel and Docker deployment.
    • Delivered a blockchain and Web3 client project in Vue.js, backed by Symfony and Node.js over GraphQL.
    • Worked directly with clients to turn requirements into shipped mobile and web features.
    FlutterDartSymfonyGraphQLVueWeb3Docker
  4. Full-Stack Developer & Project Lead at MerkleTree Foundation

    Jun 2021 to Nov 2021

    • Delivered a free STEAM education platform for Hong Kong ethnic minorities in a team of two, owning design, curriculum, marketing and project management.
    • Deployed the full AWS stack: S3 for assets, EC2 hosting Postgres and React behind PM2 and Nginx.
    ReactNode.jsPostgreSQLAWS

Education

The University of Hong Kong logo

The University of Hong Kong

BSc Information Management, Faculty of Education

2019 to 2021

Hong Kong's top-ranked university. The degree is about how information creates value in an organisation: designing the data model, governing it, and building the systems people actually use. It is why I can sit with a CRO or an editor and turn a business process into a schema and a workflow, not just a screen.

  • Data Management and Database Systems
  • Information Policy, Ethics and Digital Governance
  • Sociology of Information Systems
  • Web Development and Data Optimisation
Hong Kong Baptist University CIE logo

Hong Kong Baptist University CIE

Associate Degree in Mobile Information Technology

2017 to 2019

Mobile first, from the start. Most customers now meet a business on a phone, and this is where I learned to ship native apps end to end, which became the App Store and Play Store apps at The Asset and the cross-platform builds since.

  • Mobile application development (Android, Java)
  • Web development (React)
  • UI and UX design

What I can own end to end我可以由頭到尾負責的範疇

Product work is the visible part. Most of a company runs on the systems around it. Each of these names the project it was done on.產品只是看得見的部分,公司大半靠周邊的系統運作。以下每一項,都註明了實際做過的項目。

Hand me the whole thing and I will run it: customers, staff, payments, releases, ads, analytics, and the data your research team is waiting on.整件事交給我就可以:客戶、員工、收款、發佈、廣告、數據分析,以及研究團隊急需的數據。

  1. 01

    The whole business, not just the website整個業務系統,不只是網站

    Customer and client records, staff accounts and roles, products, documents, invoicing and the admin panel the team actually works in every day.客戶資料與紀錄、員工帳號與權限、產品、文件、發票,以及團隊每日實際使用的管理後台。

    Done on實際案例ANSR portalThe Asset

  2. 02

    Taking payments收款

    Stripe subscriptions with webhook reconciliation, PayPal checkout, and a manual bank-transfer flow with proof-of-payment upload where cards were not an option.Stripe 訂閱配合 webhook 對帳、PayPal 結帳,以及在無法使用信用卡時的人手銀行轉帳流程,並附上付款證明上傳。

    Done on實際案例DesireAITechShopANSR portal

  3. 03

    Shipping everywhere, then keeping it up發佈到所有平台,然後維持運作

    Native iOS and Android through store review on both platforms, web on Vercel and on AWS EC2 and S3 behind Nginx and PM2, containerised with Docker.原生 iOS 與 Android 通過兩邊商店審核;網站部署在 Vercel,以及 AWS EC2 與 S3 配合 Nginx 和 PM2,並以 Docker 容器化。

    Done on實際案例The AssetGeowitnessMerkleTree

  4. 04

    Selling against it靠它帶來收入

    Sponsor and advertising placement managed from the admin panel rather than hard-coded, and the reporting to show what any of it actually did.贊助與廣告位置由管理後台控制,而非寫死在程式裡,並附上報表,看清楚每一項實際帶來了甚麼。

    Done on實際案例The Asset

  5. 05

    Measuring it量度成效

    Google Analytics and Firebase instrumentation, event and funnel tracking, and error monitoring so a failure surfaces before a customer reports it.Google Analytics 與 Firebase 追蹤碼、事件與漏斗追蹤,以及錯誤監控,讓問題在客戶投訴之前先浮現。

    Done on實際案例The AssetDesireAI

  6. 06

    Getting the data out for the people who need it將數據交到需要的人手上

    SQL extraction against production shaped to what the research team asked for, plus Python, JavaScript and Laravel automation on the panel so a recurring pull stops being somebody's weekly job.按研究團隊的要求,直接在正式環境的資料庫寫 SQL 抽取數據,再以 Python、JavaScript 和 Laravel 在後台自動化,讓定期的數據抽取不再是某人每星期的例行工作。

    Done on實際案例The Asset

Skills

Every skill here is tied to a project that used it. AI and automation first, because that is where the current work is.

Projects

Hover a project to light up its skills. Hover a skill to see where it was used. Click to pin.

Skills, grouped

AI & ML

Automation

Microsoft Power Platform

Languages

Frontend

Backend

Mobile

Security

Data

Cloud & tooling

Ask this site

A small retrieval agent grounded in this page. It answers only from the case studies and roles above, cites the section it used, and says so when it has nothing.

Ask it anything the page answers. It cites the section, and it says so when it has nothing.

  1. 01Your question is matched against the site's content with BM25. No embeddings, no second vendor.
  2. 02The top matches go to a language model as numbered sources, with one rule: answer from these or decline.
  3. 03Each claim carries a citation that links to where it came from. Off-topic questions get a fixed refusal.
  4. 04With no model configured, or if it fails, the same sources are answered by picking the sentences that match. Never an error, never invented.

Questions are answered server-side and are not stored by this site. Any API key stays on the server. Rate limited.

    Tooling for AI agents

    An MCP server is how an AI assistant gets a capability it does not have. This is one I built, published and run, shown the way an agent sees it.

    Not everything worth shipping has a screen.

    hk-markets-mcp is a Model Context Protocol server I wrote so any AI assistant can reach live Hong Kong monetary data. Three tools, each with a typed input and output schema, over the Hong Kong Monetary Authority's public API. The client caches, collapses concurrent calls into one and rate-limits itself, because the thing upstream is a central bank's free service.

    Open source under MIT, 22 tests, published to npm and installable with npx. The responses shown here are real, fetched when this page was built.

    $ npx -y hk-markets-mcp

    npmSource on GitHub

    get_base_rate {}

    base_rate_percent
    4.25%
    as_of
    2026-11-20

    get_hibor {}

    overnight_percent
    2.38%
    one_month_percent
    2.71167%

    get_peg_status {}

    strong_side / weak_side
    7.75–7.85
    aggregate_balance
    HKD 54,186M

    Source: Hong Kong Monetary Authority, 2026-11-20 — its most recent trading day.

    Building next

    Work in flight, not yet shipped. Listed here so what is live and what is still being built are never confused.

    1. RegTech RAG Assistant

      In progress, Q4 2026

      Retrieval-augmented question answering over public SFC and HKMA circulars. Every answer quotes and links the paragraph it came from, and the system declines to answer when retrieval confidence is low.

      Next.jsSupabase pgvectorLLM API
    2. Agentic Event Ops

      In progress, Q4 2026

      An event-registration triage flow rebuilt from a scripted RPA job into an LLM agent, with explicit human approval gates and a decision log for every call it makes.

      n8nLangGraphPython
    3. On-device AI mobile app

      In progress, Q1 2027

      A SwiftUI app running a small language model entirely on the handset, so no text leaves the device. Measuring tokens per second and memory against real hardware.

      SwiftUIMLXCore ML
    4. Recommendation evals harness

      In progress, Q1 2027

      A fixed test set and scoring harness for the DesireAI ranking pipeline, pairing deterministic checks with a model judge so recommendation quality is measured on every change instead of eyeballed.

      TypeScriptVitestGitHub Actions

    Certifications

    • NVIDIA Generative AI and LLMs

      in progress, expected 2027

    • Claude Certified Architect

      in progress, expected 2027

    • AWS Certified Developer Associate

      planned

    • PL-900 Power Platform Fundamentals

      planned

    Get in touch聯絡我

    Open to AI engineering, full-stack and mobile roles in Hong Kong or remote, and to project work for clients. Email is fastest.歡迎洽談香港或遙距的 AI 工程、Full-stack 及手機開發職位,亦接受客戶項目。電郵最快。

    harveyworkhk@gmail.com