Answers to the most common questions about working with slonge solutions.
Swiss SMBs and mid-market companies with (a) a concrete, high-volume process or (b) a knowledge base that keeps generating the same questions. Typical indicators: substantial inbound or document volume, established data, the need for an audit trail. Poor fits: one-off tasks, very small volume, or problems that classical automation already solves cleanly — there I'll tell you straight that AI is overkill.
Three practices: Agent Systems for enterprise data migrations and workflows (SAP, JiVS, audit trail). AI Engineering for SMB applications (RAG, document extraction, support agents, multilingual triage). AI Consulting for architecture reviews and strategic guidance — when you build yourself or evaluate a vendor. Details on the respective practice pages.
Good candidate: a recurring problem, enough volume to make automation pay off, a data foundation the system can work with, and someone who reviews the output. Poor candidate: one-off tasks, tiny volumes (e.g. under 20 requests/day for triage), or problems classical rule-routing already handles. For too-small scopes, classical automation or a well-maintained Excel is often cheaper and more transparent — I'll tell you that before you invest time. See also the use-case library for concrete applications.
Default: Claude (Anthropic) for reasoning-heavy tasks and multilingual Swiss contexts (DE/FR/IT). Azure OpenAI Switzerland when data residency in Switzerland is a contractual requirement. Open-source models (Llama, Qwen, Mistral on vLLM or Ollama) when full on-premise is required — with a performance trade-off I communicate openly. Model choice is part of architecture, not a marketing decision.
Standard setup: Azure OpenAI Switzerland region (North/West). Data doesn't leave Switzerland; Microsoft does no training on your data (default off, contractually). FADP-compliant retention and logging concepts are defined per use case. For stricter compliance: on-premise variant with open-source models. EU AI Act: I plan architectures with audit trail and risk classification from the start — compliance is not bolted on.
Three pillars. (a) Eval gold set before production: 50 to 200 questions with expected answers, automated scoring on faithfulness, context precision, answer relevance. (b) Continuous monitoring in production: confidence scoring, drift detection, escalation rate. (c) Quarterly reviews with native speakers or subject-matter experts. Without eval, AI is a gut-feel project; with eval, it's an engineering project with verifiable output.
Prompt engineering for simple structured tasks (classification, extraction with a clear schema). RAG whenever the model needs access to a knowledge base (HR documents, product catalog, contracts, FAQ). Fine-tuning only when (a) training data accumulates, (b) the output format is extremely specific, or (c) latency and cost become real arguments. The classic industry mistake: everyone wants fine-tuning but actually needs better RAG. RAG first, fine-tuning later if at all.
Four phases. (1) Discovery: sharpen the use case, define pass/fail, often 1 to 2 days. (2) PoC: 4 to 8 weeks, bounded scope with eval gold set and a clear acceptance criterion. (3) Production: hardening, monitoring, handover. (4) Operation: eval reviews, maintenance or handover to an internal team. I'll tell you early if a PoC is failing — no sunk-cost continuation.
Discovery: 1 to 2 days. PoC: 4 to 8 weeks. Production hardening: 2 to 6 weeks. A first production system is realistic in 8 to 16 weeks total — depending on data quality, compliance requirements, and integration effort. Pure architecture consulting or discovery engagements are often faster.
Concrete numbers we discuss individually based on the use case — setup PoC and ongoing costs have very different drivers (document volume, model choice, eval depth, compliance level). An initial discovery call is free. ROI can't be cleanly calculated without a PoC — what I can deliver upfront is an honest range and a clear pass/fail definition.
Both. Many projects run remotely. For discovery workshops, sensitive topics, or on-site data residency setups, I'm happy to come in. Based in St. Gallen, working across Switzerland.
German and English fluently (both native-level). French well. For multilingual Swiss setups I'm the right partner — particularly because I know from experience where LLMs are strong with IT/FR/Swiss specifics and where they aren't.
Easiest is via the contact form or directly by email at kontakt@slonge.ch. I typically respond within 1 to 2 business days.
