Job-search screening system

AI agents wrote the code. I worked out the rules and checks with them in natural language and had the results blind-reviewed by a separate AI model with no preset memory.

The principle

Filter first on hard conditions. Then rank the eligible jobs by match and personal preference.

What the model does, and what it does not

Eligibility decisions stay in deterministic code. The language model ranks jobs and drafts text. When the model raises a problem with a job, it has to quote the ad. Code then looks for those exact words in the ad, and the problem counts only if they are there.

An outside reviewer

A separate AI model with no shared context reviewed the 25 top-ranked jobs. It found three real bugs. The eligibility checks were then rewritten as one set of rules instead of being patched one bug at a time.

Is it repeatable?

In repeat-run tests, disabling sampling raised identical outputs from 438 of 480 (91%) to 480 of 480 (100%). This measures consistency, not accuracy. The ranking was never tested against employers' actual responses.

Returns and claims calculatorOct 2026 – present

It turns scattered business records into plain answers about money: what was lost, and what can still be recovered.

In most businesses, those facts are spread across files that nobody has time to join up, and a wrong figure costs real money. So the tool has one rule: be right, or say what is missing.

Job-search screening systemJul – Oct 2026

It screens job ads for me in two steps. Fixed rules decide which jobs I am eligible for; a language model then ranks the rest by fit.

The job store holds over 20,000 records, far more than I could read by hand.

Agent workspacesince Aug 2026

A workspace where AI agents do my everyday work, with shared memory and written rules.

It keeps my conversations and memory in one place, instead of scattered across different AI products. A method worked out once is written down and reused, so the workspace is meant to get more useful over time.

EchoMar – Jun 2026

A prototype web app designed to help people aged 60 and over in Melbourne's city centre get out of the house and back into community life: walking routes, places to eat and drink, and nearby support services, built on the City of Melbourne's open data.

Loneliness in later life is a public-health problem, and the team found that many older people who live alone want to go out but lack a reason and the confidence. University capstone, five-person team; I was the team's Business Analyst and Design Lead.

Livestream transcript pipelineSep 2026

I wanted to build something for China's export trade, and I found a public livestream where both sides of that market met: factories and export companies looking for ways to put AI to work, and the technical people offering it. They talked for hours about real needs, pain points and what it took to deliver. I built a pipeline that recorded those talks and turned them into a searchable, time-stamped transcript, so every statement could be traced to its speaker and moment.

The room had no fixed schedule, and a session could run for more than a day. The host left the stream on even while asleep, on purpose: the host recorded and summarised the talks too, and encouraged viewers to do the same. So capture had to keep going unattended. One night the audio dropped out four times; after the fix, the next day ran 8 hours 45 minutes without a gap.

The recordings showed me two things. The businesses wanted something that works the moment they open it and makes them more efficient, the same thing I heard when I asked export business owners directly. And what was on offer was uneven: some products were thin wrappers around someone else's AI model, and some claims did not hold up. So I set a high bar for my own product: the smallest version worth offering is one people can trust, not just one that works.

How I work

I solve problems in a loop: understand what is already known, collect what is missing, pin down the problem, choose the simplest method, set a goal that can be checked, then act. The steps repeat, and new evidence strong enough to change the plan sends me back to the problem and the design.

Education and experience

My full name is Junrui Tu. I go by Tauri.

Contact