Mashwork Ltd, London
The best way to predict the future is to invent it.
Alan Kay, computer scientist
Mashwork is a London studio. We consult on generative AI, agents and Bayesian inference, and we build the iPhone and web apps that come out of that work.
What we do
We work both ends of the problem.
The useful part of AI moves every few weeks, and most teams have neither the time to track it nor the appetite to bet on it. We run the experiments, then build the product that survives contact with real users. One engagement usually turns into the other.
Consulting
Gen AI and applied research
You bring the question: whether an agent can do this job, which model to build on now that a new one has landed, why the pipeline works in the demo and not in production, what the data will actually support. You get a straight answer, the code and evals behind it, and a recommendation you can act on without us.
- LLM systems: agents, tool use, retrieval, evaluation
- Model selection and benchmarking as the frontier moves
- Bayesian inference and probabilistic modelling
- Experiment design, causal inference and uncertainty
- Technical due diligence and second opinions
Product
iPhone, Android and web apps
We build and ship the apps ourselves, end to end: native iPhone and Android, web where it fits, from first prototype to a live release on the App Store. Small team, direct contact, nobody in between.
- Native iPhone and Android apps
- Web apps and prototypes for testing an idea in days
- Model serving, evals and cost control in production
- Ongoing maintenance and support
A working demo is not a product.
Everything hard happens after the demo: evals that catch regressions, latency and cost you can live with, and the long tail of inputs nobody thought to try. That is the half we are usually called in for, so we build it too.
How we work
Two weeks to something real.
-
01
Scope
One call, then a written scope with a fixed price and a date. No month-long discovery phase.
-
02
Prototype
Something running inside two weeks. An agent doing the job, a scored eval set, or an app you can tap.
-
03
Build
Weekly builds you can use. One point of contact. Your repository from day one.
-
04
Hand over
Documentation, a walkthrough with your team, and a support arrangement only if you want one.
Gen AI and agents
Claude . GPT . Gemini . Open-weight models . Agent orchestration . Tool use and MCP . Retrieval . Eval harnesses . Fine-tuning . Inference infrastructure
Inference and experiments
Python . PyMC . Stan . NumPyro . Hierarchical models . Bayesian A/B testing . Causal inference . Uncertainty quantification . PyTorch
Product and engineering
Swift . Kotlin . React Native . TypeScript . Next.js . Node . Postgres . AWS . App Store releases . Figma