Provider-Agnostic Architecture
This site's own AI chat assistant is a live, citable example of the AI Reliability line's principles in production: a provider-agnostic streaming proxy that can switch between LLM providers through a single configuration value without any client-side change, sitting behind a deterministic, pre-prompt security-harness gate that blocks high-severity prompt-injection and system-prompt-leakage content before it ever reaches the model.
Cost Guardrails & Retention
The assistant runs under layered, independent cost guardrails — a per-visitor rate limit, a daily conversation budget that survives process restarts, and a per-session cap — so usage stays bounded and predictable rather than open-ended, and exhausting any one layer degrades to a polite static message, never an error page. Every conversation is retained for a limited period and then automatically deleted by a scheduled retention sweeper, matching this site's own published privacy notice — the deletion is enforced by running code, not just promised by a policy page. The specific threshold values are deliberately not published: they are operational tuning knobs, reviewed against real traffic, not commitments.
Scripted Evaluation
Before any prompt or model change ships, a scripted evaluation runs against the live endpoint with a golden set of prompts across three assertion classes: factual precision (answers about services must carry the exact figures published on the page), anti-fabrication (questions about services that don't exist must be declined — never given an invented price), and refusal correctness (prompt-injection attempts must be blocked by the deterministic harness before the model is ever called, while legitimate security questions — "do you test jailbreak resistance?" — must NOT be over-blocked). The full set must pass before a change is deployed, and results are recorded in a tuning journal so every model/prompt decision has a written baseline. The assistant also mirrors the visitor's language — any language, verified live in German, Polish, and Czech conversations after release.