Models, data and
software, built for your problem.
We design language models, create the data they learn from, and build the software they run in. The same people do all three, so the model and its training set are made for each other. When AI is the wrong answer, we say that first.
One team for the model,
its data and its tools.
Custom models
We design a model for your task and train it from random weights on hardware we own or rent. XE5 shows the pattern: a new skill is a new specialist, so adding one does not disturb the others. Cassandra T1 shows the open end: 1.33B parameters, weights public.
Training data
Going from 237 to 2,611 examples did nothing for our models. Adding the right kinds of examples did. We build supervised sets by construction family and anchor-masked corpora, and we have built corpora from SEC filings in the US and Canada.
Software for LLMs
Assistants that answer from your own records and cite the row they read. MCP servers with scoped tokens. Agent tooling that runs coding agents in separate worktrees. We ship our own platform, Sophia Stack X, with 53 MCP tools, and run assistants in client systems.
Corpus in,
report out.
Pick a stage. The training stage uses anchor-token masking, our method, with a DOI and 504 published generations behind it.
We train from random weights, or adapt an open model, on hardware we own or rent. Anchor-token masking keeps the tokens that matter, such as names and part numbers, intact while the rest is masked, which is the method behind our 1.67× held-out result.
A checkpoint, the training log and the configuration that made it.
Anchor tokens (orange) stay fixed. The model has to fill the masked ones (▯). Illustrative tokens
We will tell you when
it should not be AI.
Workflow audit
Before we build, we map the workflow and say where AI helps and where it does not. A process change, a dashboard or an API integration is often cheaper than a model, and we recommend it when it is.
Automation
Integrations, routing and dashboards that remove a manual step. This is ordinary software work and we do it without a model when no model is needed.
Agentic workflows
Several narrow agents with checks between them. Constellation is our own design: six lanes and an arbiter that abstains when the margin is thin and hands the decision to a person.
Four steps, in this order.
A call
You describe the task and the data. We say whether we can help and what we would need.
A written scope
What a correct answer looks like, the held-out examples we will score on, and what we will deliver. Written before any training.
The build
Data, training and evaluation as in the pipeline above. You see the failures along with the wins.
Handover
The weights, the data, the training configuration and the report are yours. We can run it for you or leave it with your team.
Client work is private. What we publish is our own research.
Describe the task.
Thomas reads it.
SOPHIA XT is the AI lab of 3 Point Forward. If the problem is a business problem first, 3 Point Forward can define it with you and we build what the plan needs.