AI engineering for public sector and enterprise
Who we have built for
How we work
What we do
01
We find where AI can genuinely change the work, check the data and constraints behind it, and turn the best opportunities into a plan you can act on.
02
We design and build complete products around the people who use them, from specialist internal tools to services for your customers.
03
We connect your documents, databases and systems so people and AI work from the same current information, with each person seeing only what they should.
We run this ourselves: the Twin's graph engine answers from live records with a certificate.
04
We build agents that gather evidence, use the tools you approve and move work forward, with clear limits and a person to hand over to.
We run this ourselves: the Twin's agent runs survive restarts and deploys.
05
We make quality and safety measurable: tests on real examples, access controls and a record of what the system did and why.
We run this ourselves: the Twin's answers are tested on agreed sets before each release.
06
We deploy into your environment, then monitor, support and improve the system as real users and changing data arrive.
We run this ourselves: one customer per dedicated server, released through one pipeline.
What we build and run ourselves
A Rust engine holds each organisation's records and agreed definitions in memory and works out every figure itself, with a certificate tracing it to its records. No AI model does the arithmetic.
Rust · semantic model · access ledger
Open-weight models served from dedicated UK or EU infrastructure, with a draft model for speed and adapters trained on real task trajectories.
Open-weight models · LoRA · speculative decoding
Each organisation's Twin runs on its own dedicated server, released through one pipeline: checks, one test run on the release, staging, then production on the same build.
Hetzner · immutable images · blue-green releases
Long agent sessions are claimed, leased and streamed, so a deploy or a dropped connection never loses or repeats the work.
Electric · durable streams · Postgres
Recurring work written as Python runs isolated, with no network and no secrets, and replays accepted results rather than redoing them.
Python · isolated sandboxes
One service holds the provider credentials, routes each call, and records what it used. A pinned model is the model that answers, or the call fails visibly.
Model gateway · routing · usage receipts
Runs, model calls and refusals are traced across every server into one store, and missing telemetry reads as unknown, never as healthy.
ClickStack · OpenTelemetry
The Twin and our fleet are operated through CLIs built for people and agents alike: JSON out, explicit writes, every change naming its organisation.
Twin CLI · Fleet CLI
A desktop client drawn natively in Rust, in development, alongside the Twin's web and iPhone apps.
Rust · native UI
How an engagement runs
Stage 1
We look at what your team does today, where time is lost and what a useful improvement would look like.
You hold at the endA clear problem, a practical first scope and a way to judge the result.
Stage 2
We map the workflow and test the approach with the people who will use it. Data, access and review requirements shape the design.
You hold at the endA first design, an engineering plan and agreed responsibilities.
Stage 3
We build the system and check its behaviour against your work. Your team can see progress and help identify what needs to change.
You hold at the endWorking software, evaluation results and a clear view of its limits.
Stage 4
We agree how the system will be supported, who owns each part and how improvements will be prioritised.
You hold at the endDocumentation, an agreed support plan and a practical improvement backlog.
Our work
Public sector · skills
Skills Compass, built for IfATE and supported for Skills England
Mercury Labs built Skills Compass for IfATE, now Skills England within DWP, between 2024 and 2026, and continues to host and support it.
Manufacturing · skills
A data platform for mapping manufacturing capability
We built the WFH Datacube for capability mapping and map-and-gap analysis, alongside OrCa Explorer, a public portal for finding organisations and capabilities by what they do.
Market intelligence
AI-powered research at industry scale
Product and engineering work for a market-intelligence platform designed to make specialist research faster and more accessible.
Financial services
Technology for retail participation in capital markets
Digital product and engineering work connecting public companies with their retail investor communities.
Health technology
A digital platform for faster access to mental-health support
Engineering a service that connects people with certified therapists, coaches, psychiatrists and wellbeing professionals.
The MTC work was funded by Innovate UK and HVM Catapult. See all our work
Where it runs, and security
Amazon Bedrock · Lambda · ECS · RDS · S3
In your own account and region, under your access controls.
Azure OpenAI · Functions · Container Apps · Azure SQL · Blob Storage
In your own account and region, under your access controls.
Vertex AI · Cloud Run · BigQuery · Cloud Storage
In your own account and region, under your access controls.
Cyber security certifications
Mercury Labs, the company behind the Company Twin, holds both certifications through its legal entity, Mercury Digital Assets Limited. Scope: whole organisation.
The product we built
It connects the AI tools your team already uses to the systems you run, so every answer works from live records, and each person sees only what they are allowed to.
Start a conversation
