The maker

Wiqi Lee

Arrearly was designed and built solo for the Build with Gemini XPRIZE hackathon - a business that operates through AI, from the decision engine to the interface you are reading.

Wiqi Lee

INTJ11 years building
Data Scientist & AI/ML ResearcherSoftware EngineerCellist
PythonJavaRustJulia

Eleven years across data science and software engineering, working where machine learning meets systems that have to hold up in production. Arrearly is that intersection made concrete: a model that predicts, an agent that decides, and an engine fast enough to re-score a whole book before anyone has finished their coffee. Away from the keyboard, a cellist, which is where the taste for structure and timing comes from.

Context

What Arrearly is, and who it is for

Arrearly is an entry in the Build with Gemini XPRIZE hackathon, whose brief was direct: build a real business that operates through AI, acquire real users, and generate real revenue inside a ninety-day window. Not a demo, and not a prototype - something that runs in production and does genuine work.

The problem

Small businesses are owed money constantly and almost none of them have anyone whose job is getting it back. The follow-up is repetitive and slightly uncomfortable, so it gets postponed, and a twenty-day-late invoice quietly becomes a ninety-day-late one halfway to a write-off.

The solution

An autonomous agent that reads each overdue invoice in full, decides the single right action for today, writes and sends the message in the right tone and channel, negotiates a plan when a customer is strained, holds when a promise is still in date, and escalates to a person only when judgement truly requires it. Every call is logged with its reasoning and confidence.

Why build it

The interesting part was never the writing, it is the judgement: whether to chase at all today, how hard, through which channel, and when pressing further costs more relationship than it recovers. That judgement is exactly the kind of work an agent can own end to end rather than merely assist with.

Who it is for

Small and mid-sized businesses that invoice and wait: design studios, agencies, clinics, trades, consultancies. Anyone with receivables and no dedicated collections person. The first live user is Heytugas.com, a real services business.

What makes it different

Existing tools automate sending a fixed reminder sequence on a timer. Arrearly automates deciding. A model proposes each action, a deterministic policy engine checks it against the escalation ladder and operating limits and steps back anything that breaks a rule, and an offline engine keeps deciding when the model is unreachable. The model reasons, the rules keep it safe, the audit trail keeps it honest.

Approach

Three ideas the build is organised around

The decision is the product

Anyone can have a model write a polite reminder. The value is in the judgement around it - whether to chase at all today, how hard, and when to stop. Arrearly treats that judgement as the thing worth building, and the message as a by-product.

Autonomy has to be auditable

An agent that acts on its own is only trustworthy if every decision can be challenged after the fact. Each run keeps the model's own reasoning next to the evidence of what was carried out, so nothing the agent does is a black box.

Guardrails belong in code

A prompt can ask a model to behave. Only code can guarantee it. The escalation ladder and the contact limits are enforced after the model answers, so no amount of drift or injection can produce a message that breaks them.

Pilot users

Voices from the businesses using it

Reviews come only from real pilot users, in their own words. Nothing here is written by the maker.

Loading…

Everything here - the agent, the policy engine, the interface - is the work of one person.