Klaus: On-Prem, Citation-Grounded Contract AI

Compliance teams reviewing vendor contracts face a trust problem, not a speed problem: 92% of contract-management errors are human error, and organizations lose an average of 8.6% of total contract spend annually to cost leakage from missed clauses. Most AI contract tools compete on drafting speed, which doesn't fix this — a fast answer you can't verify is still a liability. Klaus is an on-prem, citation-grounded RAG assistant built around one structural rule: no answer without a citation to a real document and clause. Every model citation is checked against retrieved clause metadata before it's returned; a failed citation gets one retry, then a clear refusal instead of a guess. Every query, successful or rejected, is written to an append-only audit log. he production backend serves Qwen/Qwen3-14B via vLLM on real AMD ROCm hardware (Radeon/RDNA3), verified end-to-end: captured server logs, real grounded Q&A with correct citations, and genuine throughput benchmarks (~1,125 tok/s aggregate via vLLM's own bench tooling). Qwen3 was chosen over Gemma 4 specifically because vLLM lacked native support for Gemma 4's brand-new architecture at build time — a real engineering tradeoff, documented in the repo. Running entirely on-prem isn't a technical nicety, it's what makes the product usable under real legal constraints: HIPAA Business Associate Agreements most consumer LLM APIs don't cover by default, GDPR data-residency and Standard Contractual Clause requirements, and CMMC/ITAR air-gap mandates for defense and government contractors. Zero outbound network calls in the core pipeline, by design. The repo includes hybrid dense+BM25 retrieval, clause-level chunking (not arbitrary token windows), PII/PHI detection, full test coverage with CI, and a free-hosted live preview that lets you upload your own contract and query it directly — clearly labeled as a lightweight preview distinct from the verified AMD backend.