Better AI solutions, built on how your best people actually think. Sharper with every interaction.
AotearoaNew Zealand
Generative AI SpotlightThe problem
Everyone runs the same off-the-shelf models. Your edge is the judgment your people can't write down.
Expertise isn't a document. It's the call made under pressure, the exception caught in the act.
When your expert leaves, their judgment goes too. Encode it and it stays, spreads, and sharpens.
What we do
We explore the AI solution that actually makes sense for your company. No fluff.
We capture how your best people think about the problem, then compare that judgment to the best-in-class solution.
We share the solution with your team, so you see exactly what we found and what we propose.
We review your data layer to see what can be safely used to build your custom AI.
Your data stack stops being static: a living AI that compounds every time it is used.
Getting started
No expert interviews. No months of knowledge capture. Whether you're using Confluence, Slack, or SharePoint, wherever your data is housed, the Tacit MCP (Model Context Protocol) server pulls it all together and transforms it into a pipeline our AI trains on and improves asynchronously. The output is scenarios: real problems from your daily work, each with a measured definition of what a good answer looks like, validated by your experts.
Sources we ingest
Any knowledge source
Any chat & voice
Any system of record
Any learning & ops
Connected assistants
Your specialized AI
Trained on your scenarios, validated by your experts.
You can get started with far less data than you think.
How do we know it works
Tacit works with your team to produce a specialized model, trained on your scenarios and deployed in your environment. In both runs below, the specialized model beat its frontier-scale baseline on a domain-specific benchmark.
The 0.6-billion-parameter model trained for Nourish is 390x smaller than the 235B, and beats it on calorie accuracy.
A 4-billion-parameter model trained on 106 expert scenarios beats a model 60x its size on insurance case-management reasoning.
The stack
We have four pillars that power our process: Evals, training, deployment and continuous learning. Each step powers your SLM.
Define real-world cases in the Scenario Designer and capture how your experts reason.
Train open-weight models for your use cases.
Dedicated or serverless inference for your custom models, with hot-swappable low-rank adapters (LoRA).
Capture out-of-distribution scenarios and compound model performance over time.
Our specialty
We capture the expertise no other AI company has figured out how to capture: knowledge no manual covers, built from years inside one organisation.
A senior underwriter spots a fraudulent claim in thirty seconds. We encode the judgment behind that glance.
Trust
Your AI runs inside your environment, with full control over what it learns from.
Deploys in your VPCSupports AWS · Azure · Databricks · Snowflake
You decide what data the system learns from, and which updates are approved before they go live.
Each training run is visible, auditable, and measured: what changed, why it changed, and how it performs.
Trained and deployed inside your VPC. Your data never leaves your walls.
Where we're headed
Heard at Davos
“If you're not able to embed the tacit knowledge of the firm in a set of weights in a model that you control, by definition you have no sovereignty. That means you’re leaking enterprise value to some model somewhere.”
Work with us
