What we’re building next.

Australian Jobs and AI

Forecasting AI’s effect on every Australian occupation, out to 2036.

Building it meant taking 3,036 individual work tasks through four layers - what AI can do now, how fast it actually gets deployed, the work that must stay human for legal or safety reasons, and the demand AI creates - then rolling those up into a score for all 358 Australian occupations at 2026, 2030, 2033 and 2036. Underneath all four sits the Capability Frontier: a model of how far machine capability reaches at each horizon, which is what turns a snapshot into a forecast. 12 months of development, currently under independent peer review, calibrated on Australian data: the ASC task weights, ABS Labour Force series back to 1986, and JSA occupational data. As far as we know, nothing else forecasts work this far forward.

It exists to answer one question: how much of the work people do today survives, and by when. When it launches, the scores will be public and the reports that answer what to do about them - which parts of a role change, where people still matter, what to learn now - will be the paid layer, built for associations and employers planning hiring and training.

Tasser.ai Skills Library

An interface for skills that otherwise live as files nobody opens.

A skill holds one piece of expertise as a file: the steps, the standards it has to meet, and the background facts. The AI loads it only when the task matches, so you never retype the same instructions and the context stays small. That works, but it leaves every capability buried in a folder the user never sees. The research here is the interface: how to show a person what their AI can already do and when each skill fires, without asking them to read the files.

Tasser.ai Learning Hub

Testing what builds AI capability in a workforce, not what gets completed.

Course completion is easy to measure and tells you almost nothing about whether someone can now do the work. So the research sits in the curriculum: what goes in each week, in what order, and what evidence would show it reached someone’s actual job. The platform underneath - weekly tracks, automatic progress, a leadership view of who has done what - exists to make that testable across a whole workforce. First cohort running now with our first client.

Context and Memory Systems

Memory that knows what to retrieve and what to retire.

Storing context properly is hard enough. Then the system has to retrieve the right context for the task at hand, retire a fact once a later fact contradicts it, and get human sign-off before it overwrites anything. Skip one and it confidently repeats things that stopped being true months ago.

Executive Decision Systems

A record an AI can trust, and a method for using it when the decision is hard.

Business reality arrives unstructured. These systems parse it into typed records - stakeholders, commitments, decisions, evidence - and serve them to the AI over MCP, so it reads a source of truth instead of inferring from a transcript. Every record carries its evidence trail, and the systems surface contradictions rather than quietly overwriting the older claim.

High-stakes decisions fail on the parts nobody wrote down - the option dismissed in the first ten minutes, the assumption everyone shared, the trade-off that was never priced. The method on top of the records captures those as the argument happens, so the reasoning survives the meeting and judgement stays with the executive.

SongWars

A voice agent that performs live in the game and remembers the last one.

Big Smooth, the AI MC, does not read pre-recorded lines. Big Smooth holds the state of the room - who has played, what got voted down, what happened three rounds ago - and speaks new commentary against it while the game runs. A language model writes that commentary, a synthetic voice performs it, and real-time multiplayer keeps every player’s screen in step, all inside the few seconds before the moment passes.

The memory does not reset when the game ends. Come back next week and Big Smooth remembers how you played. That is a memory-management problem: what to keep, what to compress and what to drop, so the context stays small enough to work with as the history grows.

Weather in Calendar

The calendar is the app, so nobody installs one.

Bureau of Meteorology XML goes in and RFC 5545 calendar feeds come out, so the forecast lands as events in the calendar a subscriber already opens every morning, on Google, Apple or Outlook. A geospatial query matches each subscriber to their nearest forecast location, edge functions serve the feed close to them, and Stripe runs the billing. AI built all of it - the data pipeline, the accounts, the payments - to test how much of a consumer subscription business AI can run.