We run the models ourselves, on our own machines, and everything we build reaches them through one gateway. Documents, images and time series go in; text, labels, structured JSON, verdicts and forecasts come out.
It is in production at si.overbring.com, which describes what it does, what it keeps and how access works. There is no signup form there, on purpose.
What it covers
- Documents to text. PDF, Word, PowerPoint, Excel, OpenDocument, RTF, EPUB and CSV, by URL or upload, back as Markdown. Every file is malware-scanned before anything reads it, and the fetched copy is deleted.
- Scanned pages. Pages without a text layer are rasterised at 300 dpi and read by OCR. Greek and English come with the worker by default.
- Moderation and judgement. A verdict and its categories for posts, comments, listings and uploads. Document type, business party, sentiment, quality and personal data, each as a label and a score rather than a paragraph.
- Summaries, titles, tags, structured fields. A summary within a word budget, or the fields you name pulled out of prose against a schema you set.
- Translation. A Google Translate v2-compatible endpoint: the same request and response shape, one operation, the same meters as the rest.
- Forecasts. A series and a horizon in, a quantile for each future step out, for demand, traffic or stock.
- Chat, embeddings, rerank, image generation. In the OpenAI shape, on the models your workspace was granted.
- Pipelines. Extract, classify, summarise and route a document as one run, instead of glue code between five vendors.
What is kept, and what is not
Zero retention, in plain words: nothing of yours is kept. A document, its text and the answer exist only while a run needs them — delivered once, then cleared when you collect them, or an hour after the run ends if you do not.
Your numbers are kept separately — model, task, tokens and latency, per workspace and per token — so you can bill what you use and nobody else’s traffic touches yours. The same figures are an endpoint and a live stream for your own dashboard.
Why one gateway
Every operation is a prompt and a model held as data, so a model can move without your call changing. Adding capacity for a task means adding a worker that declares it: the endpoint, the request shape and the accounting stay as they are.
A worker holds a fleet credential, which cannot call the customer API, read a workspace’s usage, or learn whose work it is doing. The gateway attributes the job, not the machine.
Who it is for
Our own products run on this gateway, and so do the projects we incubate or build with partners. A customer’s document features ship on it today.
There is no pricing page and no free tier. This is not a public service: access is by invitation, and capacity is shared with our own work. If we are already working together, the fastest route is the person you talk to. If we are not, write to isaak@overbring.com.
When it makes sense
- You are paying per token for work that does not need a frontier model.
- Your data cannot leave premises you control.
- You want a cost that does not move with usage.
- The work is documents, classifications, forecasts or translations rather than open-ended conversation.
When it does not
If what you need is the largest model available for open-ended reasoning, buy the API and keep it for that. If you want a public self-serve service with a signup form, this is not it, and it will not become it.
