Guides
Short engineering notes on how Arviser scopes LLM quality, AI security, 24-hour PoCs, 40-hour prototypes, and production gates. Orientation, not legal advice.
0. AI Security & RAG
| Guide | Focus |
|---|---|
| AI Security Review vs LLM Quality Audit | Abuse vs performance — when to book which |
| AI Security Review vs classic pentest | AI cybersecurity of LLM apps vs infra/web pentest |
| AI security risks in plain language | Business-readable risks R-01–R-27 and questions for your team |
| OWASP LLM Top 10 2026 — what we check | Category mapping for scoped reviews (not certification) |
| Indirect prompt injection | Instructions via retrieved docs, logs, or tools |
| Lethal trifecta | Untrusted input + privilege + outbound channel |
| Cross-tenant retrieval | When retrieval crosses role or tenant boundaries |
| AI Security Review | Product page · RAG track |
1. Senior AI Sprints, 24h PoC & Risk-Sharing
| Guide | Focus |
|---|---|
| AI PoC (24h) vs Prototype (40h) vs Production Pilot | When to book 24h (€2,400) vs 40h (€3,900) |
| 50% risk-shared PoC and sanitization protocol | Narrow pipeline pass marks, 50% Kill waiver, IP & synthetic data rules |
| What a 24-hour AI PoC must prove | Four numbers before you build: accuracy, refusal, p95 latency, unit cost |
| Why 10+ year senior engineers with AI beat vibe-coding | Senior + AI sprint economics vs fragile demos |
2. Evaluation Baselines, Golden Sets & RAG
| Guide | Focus |
|---|---|
| Golden set eval baseline | What most pilots skip and how to bootstrap 30–100 cases |
| RAG evaluation | Gate retrieval and generation separately |
| Refusal both ways | Measure under-refusal and over-refusal in the same suite |
| Shared LLM quality metrics | Product and risk scoreboard |
| Cost and latency gates | Budget per successful task and p95 latency next to quality |
| How we score | 1–5 scoring rubric across six audit areas |
| Glossary | Terms we use in audits and sprints |
3. Production Gates, Sampling & Package Selection
| Guide | Focus |
|---|---|
| Why a working demo is not a production gate | What breaks between demo and real traffic |
| Readiness vs Audit | Compare 1-week Readiness (€2,900) and 2-week Audit (€8,000) |
| Two-week audit timeline | Week 1 and week 2 deliverables |
| Go, hold, or no-go | Explicit decision labels tied to pass marks |
| CI gates for LLM quality | Regression harness before scale |
| Production sampling | Async live traffic sampling to catch drift |
| Kill switch before scale | Owner, thresholds, action, and logs |
| Independent of the build vendor | Why second-opinion scoring matters |
| Hiring AI-QA takes months | Sprints, audits, or dedicated specialists while you hire |
4. EU AI Act Technical Evidence (Add-on Orientation)
| Guide | Focus |
|---|---|
| EU AI Act dates | Orientation table of dates we track for counsel |
| Art. 50 marking window | Dec 2026 transparency and marking orientation |
| EU AI Act evidence is not certification | What technical evidence can and cannot claim |
RSS: /guides/rss.xml
Next: Request written scope or question · AI Sprints · Startups · Sample deliverables · Checklist · Open proof