Mercity

Building research led capabilities for product and enterprise teams.

We work with companies on problems off-the-shelf APIs cannot solve: custom training, new model architectures, optimization and private research. We handle every stage, from the training data to the evaluation. You own the weights and the code.

Where our expertise lies

We keep to a narrow set of problems and try to be the best at each of them. The three below are where most of our client work and our published research has gone over the years.

Synthetic data generation

Over the years we have generated hundreds of billions of tokens of synthetic data and trained models on it. More than half our projects need a pipeline built for that specific data, with our open‑source library Simula covering part of the work. Every dataset ships with its generation code, filters and coverage report.

Custom architecture and training

LLMs are not always the right tool, and API models cannot always be made to fit. We design and train models around the task, with modified attention and custom kernels where needed, and optimize them to run small and fast, like Qwen3-8B pruned from 36 layers to 30.

Evaluation and benchmarking

What we build is used by our clients' customers, so every project starts with evaluation. We learn the domain, build a benchmark for it, and measure every change against that baseline. We run the evaluations by hand and read the outputs ourselves.

We work in two types of engagements

Product companies come to us for the R&D that sets them apart from their competitors. Enterprises come to us to put AI into their existing systems and improve how the organization runs. Either way, you own the models, data and code we build.

For product teams

Differentiation others cannot get.

You are using the same APIs as everyone else, which means everyone you compete with offers the same features, at the same price and the same speed.
We help you take that technology in-house, so you can iterate and improve on it much faster than they can.

  • 6–14 week engagements
  • You walk away with the models and data
For enterprises

AI built into your business systems.

We help you build AI into the business systems your teams already use, so they can move faster. We optimize the business one part at a time, working alongside your engineers. That speed adds up across the organization.

  • Months-long engagements
  • Embedded plug-and-play teams

Alongside the data and weights, we open‑source our tools.

Each one started as something we needed internally and got tired of rewriting for every project. They stay maintained because we run them on our own client work.

Simula Python

Deliberate synthetic data generation. Describe the distribution you want as a spec, and Simula plans, generates and verifies a corpus against it, with a coverage report at the end.

PromptKeep Python

Prompts as first-class objects. Versioned templates with lineage in SQLite, and a drop-in OpenAI wrapper that records every run: prompt version, variables, output and usage.

Coming soon
Coming soon

Bring us the problem that does not fit the API.

If an off‑the‑shelf model already solves it, you do not need us. If it gets you most of the way and the rest is what your business depends on, we should talk.

Start a conversation