Concepts and frameworks -

Concepts and frameworks

What for? Orchestration: models, tools, and policies in reproducible flows.

Example: Incoming email → classify → pull the right prompt + context → MAIR/NK approval → send.

Outcome: Same flow, same quality — with error paths/retries.

What for? Documentation “falls out” as a by-product — from changes, approvals, and tickets.

Example: Merge + approval → short architecture note + API sketch generated; a human adds two sentences of context.

Outcome: Up-to-date, cross-linked notes with almost zero documentation overhead.

What for? A fixed reasoning grid for the AI (e.g., Problem → Hypotheses → Tests → Decision).

Example: Bug report → the AI lists 3 hypotheses, requests missing logs, evaluates tests, and records the decision.

Outcome: Fewer hallucinations, more verifiable steps.

What for? Roles, approvals, and markers — so decisions are auditable.

Example: The AI drafts a customer email → subject-matter expert (SME) reviews, Legal approves → markers: ticket ID, source, version.

Outcome: Clearly logged who approved what and why — including timestamps.

What for? A shared repository for approved prompts, texts, and vector embeddings.

Example: “Reklamationsantwort_DE_v3” (complaint reply) + FAQ CSV are stored and tagged → the assistant pulls the context via tags faq,de.

Outcome: Same inputs, same quality — reusable across the whole team.

What for? One source of truth for claims and numbers — across email, deck, and one-pager.

Example: Revenue updated from 3.2m to 3.6m → all outputs (email, deck, website snippet) update automatically.

Outcome: No contradictions between documents.

What for? The AI mirrors back what it understood — before it acts.

Example: Ticket “Customer cancellation” → AI: “I see reason X; I’m missing the contract term. Should I do A or B?”

Outcome: Misunderstandings are resolved before any output.

What for? Policy check at input/output: data protection, redlines, masking.

Example: The reply contains an IBAN + name → the IBAN is redacted; the output is logged with the reason.

Outcome: Policy-compliant answers, including traceable rejections.

What for? Consolidate feeds (prices, news, files), validate them, and normalize to a common schema.

Example: Two price APIs + the German Federal Gazette (Bundesanzeiger) → map to ISIN, flag outliers > 30%, cache the result.

Outcome: Reliable, versioned data instead of API chaos.