Private deployment
When nothing may leave your network
Hospitals, banks, law firms, government, industrial R&D — teams whose policy forbids any third-party processor. CompresLM runs at three levels of isolation; pick the one your policy demands.
Nothing stored
Your prompt passes through memory, is forwarded on your own key, and is gone. What we keep is token counts — see the Privacy Policy.
The gateway on your hardware
A licensed install in your VPC or server room. Compression and the savings stay; we never see a byte.
The model inside too
We install an open-weights LLM, tune inference on your hardware and put the gateway in front. Nothing leaves your network — not to a provider, not to us.
The closed loop, in practice
What a full private install includes:
- Model choice and hardware sizing — an open-weights model (Llama, Qwen, Mistral families) matched to your workload. We help pick the smallest model that answers your tasks well, not the biggest that fits the budget.
- An inference server tuned to your traffic — batching, context limits and quantisation set for your documents, not for a demo.
- The CompresLM gateway in front — on your own hardware compression pays twice: smaller prompts mean more requests per GPU, not just a smaller bill.
- A licence, not a subscription to us — installed under a licence agreement, terms and support individual.
Data protection for LLM traffic
An optional layer on any level — asked for most where the data is most sensitive. Three parts, in plain words:
- Masking at the source. Before text leaves your secure zone, a connector replaces names, numbers and rare identifying words with keyed stand-ins — like lending a contract with the names taped over, where only your key lifts the tape. The model answers using the stand-ins; the reply is translated back automatically. The key never leaves you.
- Encryption in between. TLS on every hop, AES for anything at rest. Standard — stated so the story has no gaps.
- Confidential computing, optional. Where GPUs are rented rather than owned, inference can run inside an attested enclave (TEE), so even the machine's operator cannot read its memory.
The honest boundary: masking is pseudonymisation, not whole-text encryption — common words stay readable so the model can still reason. What it protects is identity: names, ids, amounts. We state where this class of protection ends because that is what makes the rest of it credible.
Tell us about your perimeter
Closed network, regulated data, an unusual stack — describe it and we come back with a concrete shape and terms.
Talk to us
- Connect it — we'll get it running against your traffic and check the numbers with you.
- Fit it to you — closed network, unusual documents, a shape we don't handle yet.
- Build it — no product to connect yet? Our partners build it, we make it cheap to run.
Usually a reply the same working day.