01 AI agents
An agent that does the task
The agent plans, uses your documents and tools, and remembers context. Sensitive actions wait for your “Approve”.
- You send
- A description of one repetitive task and the sample documents it relies on.
- You get
- An agent that works through the task step by step and asks for approval before every sensitive action.
- Example
- Drafting a reply to a ticket from your documentation, approved before it goes out.
See the loop from the inside →02 Knowledge bases
Answers with a footnote to the source
You ask a question, the answer is built from your documents, and every sentence has a footnote to the passage it comes from.
- You send
- A handful to a few dozen sample documents: contracts, procedures, manuals, PDFs.
- You get
- A knowledge base with hybrid search and reranking that points to the document and the paragraph.
- Example
- A law firm asking about clauses in case files that can't go to the cloud.
How we find the paragraph →03 AI infrastructure
A model measured on your workload
We run a model from the open-source family you choose on NVIDIA Blackwell GPUs and measure it on your queries.
- You send
- Sample queries and a description of the traffic: how long the queries are and how many users ask at once.
- You get
- A model served with vLLM behind one API gateway, plus measurements of response time and throughput.
- Example
- Deciding whether to set up your own GPU server or use ours.
What happens on the GPU →04 Models & fine-tuning
A classifier trained on your examples
We fine-tune an open-source model with a LoRA adapter on your labelled examples and check it on examples it hasn't seen.
- You send
- Labelled examples, e.g. customer queries assigned to categories.
- You get
- A small model that knows your domain and sorts new examples into your categories, with a score on a test set.
- Example
- Sorting queries by type, like our classifier for 11 query types.
What happens inside the model →05 R&D
A prototype on sensor or 3D data
We combine hardware with AI. On your sample we build a prototype that shows what can be drawn from that data.
- You send
- A data sample: sensor signal recordings, footage or 3D scans.
- You get
- A working analysis or visualisation prototype on your sample.
- Example
- Analysing signals from sleep studies, or an interior in 3D to explore in the browser.
Step into the workshop →