2.01
Risks to British Business
An LLM pipeline reading every UK annual report and pulling out confirmed risk events, enriched with company data and news, written to a live map. Successor to GERM.
Founded
2020
Autonomy Data Unit // Field Manual
Doc. ADU-001 Rev. 2026.06 Section 0
Public-interest data science · operating since 2020
DATA SCIENCE,
POINTED AT POWER.
Six data scientists and machine-learning engineers inside the Autonomy Institute. We scrape, model and read at supercomputer scale for unions, charities, newsrooms and campaigners. Think of the data arm of the public interest, built with the methods the other side keeps to itself.
A unit of the Autonomy Institute
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
0.1 Abstract
What this unit is for
We started in 2020 with the Jobs at Risk Index, a Covid-era map of which UK jobs were exposed first. The brief has not changed since: take a question that matters to people without much power, point real computation at it, and publish something they can use. Some of us could be at Palantir. We are not.
1.0 Capabilities
Four instruments, one bench
Each capability is documented below as a spec sheet: what it does, the inputs it takes, the methods it runs, and a shipped example you can open.
CAP-01
Network analysis
ref. six-degrees-of-reform
Function
Map who is connected to whom. Filings, donations, contracts and the open web, pulled together into one graph.
Inputs
Companies House, Electoral Commission, procurement records, scraped pages.
Method
Scrape, then LLM entity and link extraction, then graph construction and clustering.
Output
Searchable connection maps and named-entity networks.
CAP-02
Economic & labour modelling
ref. jrf-landlord-profit
Function
Model how money and work move. Bespoke indices, input-output models, microsimulation of policy.
Inputs
Administrative statistics, survey microdata, job adverts, official accounts.
Method
Microsimulation, input-output analysis, custom index construction.
Output
Headline figures and dashboards that hold up to scrutiny.
CAP-03
Document intelligence
ref. rtbb
Function
Read more documents than a person ever could. LLM extraction across millions of pages into one queryable dataset.
Inputs
Annual reports, regulatory filings, 900-page policy plans, PDFs by the crate.
Method
LLM pipelines with human verification, run on the Isambard supercomputer when the volume demands it.
Output
Structured tables and live feeds built from prose nobody else has read.
CAP-04
Tools & data products
ref. care-visa-sponsorship-database
Function
Hand the result back. Searchable databases, trackers and indexes that anyone can use in a browser.
Inputs
Outputs of CAP-01 through CAP-03, plus a partner who needs them in public.
Method
Pipelines feeding hosted web apps, kept current after launch.
Output
Public sites that outlive the report they came from.
+-----------+ +-----------------+ +------------------+ | SOURCES | ---> | CAP-01 NETWORK | -. | | | filings | +-----------------+ \ | | | donations| ---> | CAP-02 MODEL | ---> | CAP-04 TOOLS | ---> PUBLIC | contracts| +-----------------+ / | databases | | open web | ---> | CAP-03 DOCS | -' | trackers | +-----------+ +-----------------+ +------------------+
Autonomy Data Unit // Field Manual
Doc. ADU-001 Section 2 Project Index
2.0 Project Index
Selected work, with links
Sorted by recency
Live link where one exists
2.02
Labour, the Party of Capital?
Tracing Labour's shift toward business donors between 2019 and 2024, built on a longitudinal dataset of UK political donations.
2.03
The Authoritarian Stack
Millions of pages scraped to map the modern far right and its links to power. The data analysis behind a thesis that ran in Le Monde diplomatique and beyond.
2.04
AI Exposure Index · 30M Job Adverts
Thirty million job adverts tagged with LLMs on the Isambard supercomputer, measuring AI exposure across the UK economy. Built with the UK AI Security Institute.
2.05
Givers & Takers
The donor-contractor nexus at the heart of government: political donations linked to the companies winning state contracts.
2.06
The Property Premium
An economic model of landlord returns in England, built for the Joseph Rowntree Foundation. What renting actually pays the people who own the homes.
2.07
Corporate Underminers
A co-mention network built for the International Trade Union Confederation, surfacing the companies that turn up together when labour rights come under pressure.
2.08
Project 2025 Index
An AI-augmented index of the Heritage Foundation's 900-page plan, so a reader can find what it says about any given subject in seconds.
2.09
Care Visa Sponsorship Database
A searchable database of UK care providers licensed to sponsor migrant workers, built with the Bureau of Investigative Journalism to help workers find legitimate sponsors.
2.10
Arts Funding Tracker
Arts-council funding by constituency since 2014, built for Equity. Pick a seat, see what its arts budget has done.
2.11
Six Degrees of Reform
Mapping the corporate connections of the UK's entrepreneurial far right, one company filing at a time.
2.12
Jobs at Risk Index
The origin project. The UK workforce scored by Covid exposure in early 2020, picked up on Peston and elsewhere. Everything since traces back to this.
3.0 Personnel
Six people, named
A small team inside the Autonomy Institute. Small enough that the person who built the thing is the person you talk to.
ID
Name
Beat
01
Lukas Kikuchi
Lead. ML and data engineering.
02
Bhargav Srinivasa Desikan
Data science, ML research.
03
Sean Greaves
NLP, network analysis.
04
Sonia Balagopalan
Investigations, political data.
05
Luiz Garcia
Economic modelling.
06
Jeremy Kwok
Forecasting, statistics.
4.0 Partners
Who we work with
AI Security Institute
Joseph Rowntree Foundation
Joseph Rowntree Reform Trust
Equity
Unite the Union
ITUC
Good Law Project
Bureau of Investigative Journalism
Centre for Investigative Journalism
Future Economy Scotland
Spotlight on Corruption
TfL / GLA
5.0 Contact
Start a conversation
Have a question that needs real computation?
Most of our work starts with someone we know mentioning a problem. If you have data that will not sit still, a filing nobody has read, or a number you need to defend in public, write to us.
adu@autonomy.work