Applied AI Solutions Builder

Eric Leung

I'm Eric. I build practical AI tools for everyday problems.

My projects explore helping people find trustworthy information, review AI-generated code, and automate repetitive work while keeping important decisions in their hands.

I care about making these tools useful, understandable, and clear about where human judgment matters.

Preview01–13

Projects

Tools for finding information, reviewing code, and simplifying everyday work.

Selected work / 001

Three projects.
One practical approach.

These projects explore how AI can help with finding reliable information, reviewing code, and updating spreadsheets — while keeping important decisions open to human review.

Choose a project to see the problem, watch its recorded workflow, and explore the decisions and evidence behind it.

Play Danish Immigration RAG workflow
Actual execution recording — type “What is PD3?”, get a cited answer from signed knowledge release kr-2026-09-05.1 and local gemma4:12b, then open a claim’s evidence. The 18-second model wait is shortened to 3 seconds and labelled. Not legal advice.
01 Trustworthy source-grounded retrieval
Danish Immigration RAG
Find answers about Danish immigration with the source passages beside them.
What I built
A small web app that runs on your own computer. It finds passages in five official Danish sources, writes an answer with a local AI model, and shows the source for each claim.
My contribution
I owned the user problem, scope, acceptance criteria, product boundaries, and test runs. AI coding agents helped draft implementation, tests, documentation, and development plans.
Important limit
Its knowledge is a September 2026 demo snapshot and is not kept current, so later changes to the official sources are not reflected. This is not legal advice or an eligibility decision.
Case study: Danish Immigration RAG Code: Danish Immigration RAG Reviewed repository snapshot
Play Alfredo terminal run: goal, repair and review branch
Actual execution recording — real run, not a mock-up: qwen2.5-coder:14b on a throwaway sample repository. One check fails and autopilot repairs it. Two model waits are cut and labelled (11 s and 4 s); the goal appears at once rather than being typed. 25 s.
02 A terminal app for local AI coding agents
Alfredo
Give an AI model on your own computer a goal. Get back a Git branch to review; autopilot never moves your own branch.
What I built
A terminal program that runs on your own computer. You type /go and a goal. It splits the goal into small tasks and has an AI model do each one in its own separate copy of your project. It runs a check for each task, repairs failures a few times, and leaves one new Git branch for you to review.
My contribution
I decided what the program may do alone and what stays with a person: a task is accepted when its approved check passes, and autopilot never changes your own branch. AI coding assistants helped write the code, so I do not claim I wrote it all by hand.
Important limit
It runs only on Linux, and a weak check means weak acceptance. It is an early version (0.1.0) with no published release, and the recording is one run, not a measure of how often it works.
Case study: Alfredo Code: Alfredo Reviewed repository snapshot
Play Personal Wealth Tracker monthly run and commit
Actual execution recording — a real CLI run on synthetic data: setup, monthly run, Exception Sheet decision, re-run commits the month. The terminal and sheet are drawn from the run’s output files, not screen-captured; the sheet decision is scripted and Category Memory starts with one seeded prior decision. Local model waits of 15 s and 13 s are cut and labelled.
03 Automation where a person keeps the decision
Personal Wealth Tracker
Turns your monthly bank statement into an updated Excel tracker, and asks you only about the rows it is unsure about.
What I built
A small program you run on your own computer with one command each month. It can first do a practice run that changes nothing. Rules you set (the trust policy) decide, row by row, which transactions it may sort on its own, and two AI models on your computer can help suggest categories. Anything it is unsure about goes into one spreadsheet, the Exception Sheet, for you to decide. It then saves the whole month into your tracker at once, or not at all. That is your own spreadsheet, not a duplicate, and a backup stamped with the date and time is made first.
My contribution
I decided which rows the program may sort on its own and which must wait for a person. AI coding assistants helped write the code and tests, so I do not claim I wrote it all by hand.
Important limit
Shown on one month of made-up (synthetic) data only. This is not financial advice, and it does not claim the AI models are accurate or have been tried on real statements.
Case study: Personal Wealth Tracker Code: Personal Wealth Tracker Reviewed repository snapshot

Hobbies

Outside of work, I enjoy capturing everyday moments with my Nikon F3, experimenting in the kitchen, and cooking for the people I care about. I also love the energy and social side of team sports, especially basketball and volleyball.

Film Photography
I started this hobby over ten years ago. Here are some of my favorites to share.
1. 01 Man in Copenhagen
2. 02 Icelandic horse
3. 03 My friend Kellas
4. 04 My grandma
5. 05 Mongkok, Hong Kong
6. 06 Sham Shui Po, Hong Kong
7. 07 Guitar man in Hong Kong country park
Cooking
I mainly make Italian and Asian food, sometimes combining Chinese and Western styles.
1. 08 Homemade dumpling
2. 09 Carbonara
3. 10 Pan fried Hakka stuffed tofu
4. 11 Malay Curry Chicken
5. 12 Chinese egg noodles with spicy minced pork
6. 13 Homemade Tonkotsu Ramen
Watching old and new films