Alessandro Conway
I study how labor markets work, and I build the systems that measure them.
Head of data science and engineering at Radius Intelligence, now advisory. PhD candidate in economics, Université Paris 8. MPhil in economics, Sciences Po, as a Fulbright scholar.
About
I am an economist, and I work across the whole arc of a problem, from framing the research question and building the data to modeling it and turning the result into a product people use. I have done each of these stages on its own, and I have also run all of them together.
At Radius Intelligence I built two digital twins of the labor market, live models with detailed forecasting. The first shows an employer its workforce gaps and needs by occupation and facility, and the second carries that detail to a whole state, so that training providers, community colleges, employers and policymakers can see the pipelines, on-ramps and skills at risk and plan around them. Under both sits a data layer I built that reconciles 220 sources, along with an LLM system that maps any job title to the federal taxonomy.
My research spans labor economics, macroeconomics, economic development and, most recently, AI and skills, with a paper published in Oxford’s Social Macroeconomics series and a growing set of working papers. Before Radius I spent three years at Jobs for the Future, where I led the work behind Verizon’s initiative to train and place 500,000 underrepresented people in quality careers, choosing the occupations and shaping the training with a team of three, and led one of the first national studies of AI job displacement. I now also run AI at Work, a public index of AI use, jobs and training that updates with each data release.
I work mainly in Python, for product development and forecasting, and SQL, for databases on Supabase. I have also worked extensively in Julia for macroeconomic models, R for econometrics, and interactive dashboards built both in code and in Domo and Tableau.
Outside work I build and play guitars, mix songs, and make olive oil.
Work
Roles
Research
My papers and research notes are below, and the lines join pieces that are interconnected.
Research areas
Throughout my work I kept being asked about particular occupations, how many people leave them, where their workers come from and what training into them leads to, and the official statistics could answer only for broad groups. So I started building the answers myself from public data, the same way for every detailed occupation and, where I could, every state. Each paper below is one of those estimates, checked against the official figures and put to work on a question someone actually asked me.
Every question I was asked about AI and jobs came with an exposure index attached, one that ranks jobs without saying what it measures. Once the labor market is written as one wage bill at three levels (occupations, tasks and skills), it becomes possible to see what each measure can identify, and to follow AI into what employers ask for. Next come the activities AI could do but has not yet entered. The measures also feed AI at Work, a public index that updates with each data release.
I wanted to understand why Europe’s lagging regions stay behind after thirty years of cohesion funding. My PhD treats a region as an ecosystem of place, people and economy within its national system, tests that view across the euro crisis and the pandemic, and then fits a model of its dynamics to 210 regions. From here I plan causal designs for the ties between the three domains, and a network between regions to sit alongside the one within each of them.
Most UK tax changes are now announced months before they take effect, so the tax shocks needed better dating, and with that dating I could show that the same tax rise costs roughly twice as much in consumption when the economy has slack. Alongside it sits the wellbeing model from my master’s thesis, which brings a household’s social life within macroeconomics and is now being extended to France and to Europe’s regions.
Finished 2024, updated 2026 Download
Most labor data starts as messy text (job titles, job ads, reports), and I needed reliable ways to turn it into occupations and signals. The notes here cover autocoding titles to the federal taxonomy, a taxonomy of roles finer than the federal codes, and labor sentiment read from the Federal Reserve’s Beige Book, and each of them could grow into a public tool.
Systems and tools
AI at WorkEconomics, AI, live
Is AI changing American work yet? No single dataset can say, so AI at Work reads public sources side by side in three indices, on use, jobs and training, which update as new releases come out.
Use measures how much of the country’s paid work involves tasks people already bring to AI, from the Anthropic Economic Index and OpenAI Signals set against BLS wages. Jobs asks whether pay, the mix of jobs and what job ads ask for have changed more than in ordinary years. Training follows what students choose to study and whether young workers still get the jobs most exposed to AI. In the latest build, tasks people use Claude for make up 23 percent of US wages, pay has not moved more than usual, and job ads turned away from tasks AI can do between 2020 and 2022, before ChatGPT, with no further drop at its release.
Open Inequality AtlasOpen data, live
Inequality is usually quoted as a single number, the income Gini, largely because the World Bank publishes it most often. It measures what people are paid, but whether a family can buy a house or get through a lost job depends more on what they own.
The Atlas puts wealth next to income, and the two rankings often disagree. A country can have one of the most even income distributions in the world and a more uneven wealth distribution than Britain. Wealth data is harder to put together. The series combines the World Inequality Database, the ECB household survey, the Luxembourg Wealth Study, the US Survey of Consumer Finances and the Federal Reserve’s Distributional Financial Accounts, and each row is tagged by source priority and comparability tier. The Atlas also covers income, life expectancy, life satisfaction and poverty, and a US layer goes down to the commuting zone.
NED dashboard, California regional analysisEconomics, policy, live
An interactive assessment of economic development across California's regions, built on indicators drawn from what mayors, community colleges, workforce boards and employers see on the ground.
It applies the three-pillar framework from my dissertation to California, starting from what people in each region say decides whether a household can get ahead and stay there, and then checking which of those factors can be measured well enough to compare regions.
Wellbeing and macroeconomics explorerMacroeconomics, live, QuantEcon
An interactive version of the SAGE model from my Oxford paper. You can change the parameters and see how the social psychology terms affect what happens to a household during a shock.
The model will soon be available on QuantEcon, the open source project for quantitative economics, computational lectures and tools in Python and Julia founded by Thomas Sargent and John Stachurski and used widely in teaching and research.
I built these at Radius Intelligence, a workforce planning startup, in 2025 and 2026. The two digital twins are the products, and everything else here is a piece they need, all of it running on a data lake of 220 sources and more than eight million observations. Their supply and demand models were the basis of two annual healthcare workforce analyses Radius produced for a state client. The code and data are private, so each page describes the approach with figures rebuilt from synthetic data.
Employer digital twinData science, econometrics, product
A hospital or care provider wants to know whether its turnover is unusual, what is coming, and what to do about it, and its HR export does not say.
I built it on anonymized HR records from hundreds of healthcare employers. It takes an employer’s export through to a forecast of turnover and hiring by occupation and facility, and lets the employer test scenarios against that forecast.
Labor market digital twinEconomics, policy, data science
A state planning its healthcare workforce needs to know, occupation by occupation, how many people it will need and where they will come from, and the official projections are years old by the time they arrive.
It covers 17 core healthcare occupations and the training providers for them across the US, joined from IPEDS, RAPIDS, CareerOneStop and state licensing boards. Each occupation is modeled as stocks and flows that have to balance, demand is estimated four ways, and the twin reports the gap between supply and demand. Its models were the basis of two annual healthcare workforce analyses for a state client.
Occupation resolverMachine learning, LLM systems, data engineering
Employers name the same job a hundred ways, and until those titles are mapped to one taxonomy there is no way to benchmark, forecast or compare across employers.
The resolver handles any title through a ten-stage pipeline with three model layers, and both digital twins depend on it. On its first full run it took about 50,000 unique job titles from employer HR records and resolved two thirds of them, and because it keeps what it learns, a second run over the same titles resolved about four in five from memory alone.
Data lakeData engineering, data science
Connecting labor market datasets means mapping each one to the federal taxonomy, and the sources differ in units, level of detail, vintage and license.
The lake reconciles 220 of them, more than eight million observations, to one occupational and geographic taxonomy, so any source joins to any other. A shared set of definitions settles the differences in one place, and a librarian layer answers questions about what the lake holds.
Workforce blueprintsEconomics, data science
When a new facility is planned, the region needs to know how many workers it will take, in which occupations, and on what schedule.
Announcements give a headline headcount and staffing statistics describe past hiring, so the blueprint has to be built from the ground up.
Occupational opportunity scoringPolicy, economics
A workforce board mostly wants to know which occupations to invest in and what would help them grow.
It scores every occupation on five published pillars, gives each estimate a confidence band, and suggests levers only where the evidence supports them.
Task-based labor marketEconomics, machine learning
The main measures of AI exposure disagree with each other, and claims about AI’s effect on jobs have gotten ahead of what anyone can measure.
It is a calibrated production function where every parameter is sourced and can be changed. A research paper on it is in progress.
Market signalsEconomics, data science
Official statistics come out quarters late, and decisions often have to be made before then.
It grades fifteen kinds of early evidence, including job postings, layoff notices and unusual wage moves, by how much each can be trusted, and combines them into two flags, each listing the evidence behind it.
Extra problems
These are smaller pieces from the same platform, each covering a separate data problem, the figure it produces and how far it got. Read them.
CREST, green workforce readiness dashboardJobs for the Future
A national dashboard of green workforce opportunity and need for every US county, which I designed and built from scratch.
It groups counties by their mix of climate risk and economic resilience and identifies the strategies that have worked for each mix. It became the flagship product of a large national climate initiative, was used by workforce boards and mayors across the country, and was cited in a national press release.
Tech careers workforce development programJobs for the Future
I led the work behind Verizon’s initiative to train and place 500,000 underrepresented people in quality careers, managing a team of three.
I chose the six target occupations the 500,000 would train for, using demand mapping and job-quality scoring, and then helped design the training curricula around them. Those six occupations are the basis for the initiative’s national enrollment and placement.
Publications and essays
Education
Contact
Best reached by email or LinkedIn, and I am happy to talk through any of the above.
Here are some inspirations of mine: