Private AI workloads
The AI work that fits the machines you already own.
kibbu suits AI work that touches sensitive data and does not need an answer in the next second: processing documents, building embeddings for internal search, running background agents and evals, and batch jobs. That work runs on open models on your own machines rather than on a cloud provider's.
Workloads
What teams run, and how to pilot it.
Each row is a pilot. Run one, read the measure column off it, then decide on the next.
| Workload | For example | First pilot | Measure |
|---|---|---|---|
| Document processing | Extracting fields from contracts, invoices or claims; summarizing case files; tagging and classifying records. | One repeatable batch of documents with a reviewed answer set. | Share of accepted outputs, completion time, retries, deadline met. |
| Embeddings and search | Embedding and re-indexing an internal knowledge base, tickets or code for retrieval. | One corpus with a retrieval test set. | Retrieval quality on your test set, indexing time, failures. |
| Background agents | Scheduled agents that triage, draft or reconcile, with many model calls and nobody waiting on any one of them. | One scheduled agent with clear acceptance criteria. | Successful runs, completion time, failure and retry rate, output quality. |
| Evals | Running a fixed test set against open models in CI or before a release. | One fixed suite on a supported open model. | Completed suites, reproducibility, completion time. |
| Batch generation | Drafting, translating or rewriting large sets of internal text overnight or between meetings. | One recurring job with a review step. | Accepted outputs, throughput, completion time. |
Who it suits
Teams with data they can't send out.
The common thread is data that has to stay inside the company, and work that can wait a little.
- Legal and complianceContract review, policy checks and case summaries on privileged documents.
- Finance and operationsInvoice and claims extraction, reconciliation agents, internal reporting.
- Engineering and dataCode and ticket embeddings, evals in CI, data labelling and enrichment.
- Support and successTicket triage and tagging, and drafting answers from internal knowledge.
- HR and people teamsSummaries and classification over employee records that stay internal.
- IT and securityOne governed place for teams to run AI, with policies on which models and machines they use.
Honest boundaries
When kibbu is not the right tool.
Better to know before a pilot than halfway through one.
- Someone is waiting on the first token, as in live chat in front of a customer. Machines yield to their owners, so response times vary.
- The work needs a frontier model that only runs on datacenter GPUs. A laptop is a laptop.
- The work depends on a proprietary cloud model's exact behaviour. A different model gives different output, so it is a new evaluation, not a swap.
- Your machines are never idle or never plugged in. kibbu yields to the owner by design, so a fleet in constant use has little to offer.
Tools you already use
Anything that speaks the OpenAI API.
kibbu is an OpenAI-compatible endpoint, so tools connect through their existing OpenAI settings.
| Tool | The change |
|---|---|
| n8n | Edit the OpenAI credential: base URL, kibbu key, and the model on each node. |
| Dify | Add kibbu as an OpenAI-compatible model provider, then choose it per app. |
| Langflow | Set the base URL, key and model on the OpenAI component. |
| Databricks | Point the OpenAI client in a notebook or job at the kibbu base URL. |
| Your own code | Any OpenAI SDK. See the API section. |
Check each tool against the endpoints kibbu serves before you rely on it. Tools that need images, audio or fine-tuning will not work, because kibbu does not serve those endpoints.
Keep reading
Start with one workflow.
Tell us about the work you want to keep private, and we will scope a pilot on your own machines.