NexusMind Inference v1.0.0

NexusMind Inference v1.0.0

Evidence-first FastAPI inference gateway for guarded customer-service LLM and financial RAG workloads.

Highlights

• Canonical /api/deploy/ask gateway with routing, cache, rate limit, timeout/OOM fallback, metrics, citations, and guardrails.
• Frozen Windows 8-token model benchmark: 32-concurrency full P95 10.742 s; model execution P95 382.759 ms; queueing is the bottleneck.
• Frozen customer_cache run: 32-concurrency P95 27.318 ms and 1012.745 QPS; this is cache-path evidence, not generation latency.
• WSL2/vLLM Base 5/5 smoke plus one separately bounded rank-16 LoRA serving request.
• Dynamic INT8 smoke: 1.0629 s to 0.4620 s (2.301x), with changed output and no RSS reduction.
• 123 frozen core contracts plus 10 public-release audits; current combined run is 133/133.
• Responsive GitHub Pages portfolio and local safe FastAPI workbench.

Evidence boundary

This release publishes aggregate metrics, frozen identifiers, source, tests, and contracts. It excludes model weights, private RAG data, raw holdout prompts/answers, per-request benchmark JSON, and server logs. It does not claim production SLA, model-quality improvement, high-concurrency production readiness, or AWQ/GPTQ measurements.

Assets

The attached wheel is a source/metadata and utility-module distribution; it does not bundle model weights or Project 1/2 data. The sdist is the matching source package.