Kelsus

Lending and financial services teams that want their documents and decisions kept inside their own infrastructure are the clients Kelsus builds systems for. The company deploys open-weight language models on cloud accounts that the client owns, so the client keeps control of its data throughout. Its clients work with large volumes of regulated documents, and the work is built around those documents rather than around a general chat product.

Document workflows that stay in the client's cloud

The company offers three services. The first is an open-source model evaluation, in which the team writes a short memo on which open-weight model fits a defined workload, with test results, harness code and a cost projection. The second is a retrieval platform set up within the client's own cloud account, delivered with Terraform modules, dashboards, a runbook, an evaluation harness and a transfer of knowledge and ownership to the client's staff. The third is a document workflow pilot covering extraction, classification and agentic routing for a real process, using whichever model suits the job on price and evaluation results.

Benchmarks published every quarter

Kelsus publishes a quarterly index of open-weight models on the kinds of workloads that matter to financial services processes, measuring quality, speed and cost. The index is reproducible and uses synthetic data rather than customer data. It sits alongside a reference architecture for a retrieval-augmented deployment on Kubernetes, which the company has released on GitHub so engineers can read the design and its service levels. The company says it runs open-weight models with GPU acceleration in production on AWS, with Azure and Google Cloud also supported.

When self-hosting is the wrong answer

The company is open about trade-offs. Where a managed option inside a client's virtual private cloud fits, such as Bedrock, Azure OpenAI or Vertex AI, Kelsus recommends that route, and it reserves full self-hosting for cases where sovereignty requires it. Because it is not tied to one model, it has no stake in promoting a particular one. Its compliance position covers SOC 2 Type II, which is in progress, controls aligned with GLBA and FINRA rules, and HIPAA controls where medical data is in scope.

The first step with a prospective client is a short fit check, in which the team asks about the workload and the constraint and says whether open-weight self-hosting is the right answer. If it is not, the team says what is.