AI Engineering
Building with LLMs: integration, RAG, evaluation, cost control, and product patterns.
LLM Integration in SaaS: Architecture Patterns That Survive Production
How to wire an LLM into a production SaaS without runaway cost or latency. Streaming, caching, fallback, and workspace isolation patterns from real products.
Retrieval-Augmented Generation (RAG): A Practical Engineering Guide
RAG beyond the hello-world. Chunking strategy, embedding choice, hybrid search, re-ranking, and the evaluation harness that tells you whether it actually works.
AI Cost Management: Keeping LLM Spend Predictable in SaaS
Token economics for product teams. Prompt budgets, caching, model tiering, and the per-workspace metering that keeps an AI feature from eating your margin.
Prompt Engineering for Products (Not Demos)
Production prompting is software. Versioning, testing, structured output, injection defence, and why your system prompt belongs in source control with a changelog.
Evaluating LLM Features: Testing What Doesn't Have a Right Answer
You cannot ship AI you cannot measure. Building eval sets, LLM-as-judge, regression gates in CI, and human review loops for product-grade quality.
pgvector vs. Dedicated Vector Databases: Choosing for SaaS Scale
When Postgres pgvector is enough and when you need Pinecone, Qdrant, or Weaviate. Index types, scaling limits, and the operational cost of a second datastore.
AI Agents in a SaaS Product: Where They Help and Where They Hurt
Agentic features beyond the hype. Tool-use design, the autonomy spectrum, human-in-the-loop checkpoints, and why most useful agents are narrow, not general.
Claude API vs. OpenAI for EU SaaS: Data Residency and Practical Tradeoffs
Choosing an LLM provider for a DACH product: data processing terms, EU residency, model strengths, and the abstraction layer that lets you switch without a rewrite.
Designing AI Feature UX: Trust, Latency, and the Empty State
Good AI UX manages uncertainty. Streaming feedback, confidence cues, editable output, graceful failure, and onboarding users who have never prompted anything.
AI Features and DSGVO: Lawful Processing When an LLM Touches Personal Data
The compliance layer under every AI feature. Legal basis, Article 22 automated decisions, subprocessor disclosure, and keeping personal data out of training.
Fine-Tuning vs. Prompting vs. RAG: Choosing the Right Tool
Three ways to make an LLM do your task. When prompting is enough, when RAG grounds it, when fine-tuning earns its cost — and why most teams reach for the wrong one.
EU AI Act and General-Purpose AI: What Builders on LLMs Must Know
GPAI rules took effect August 2025. What downstream builders inherit, transparency and copyright obligations, and where the line sits between provider and deployer.
AI Observability: Logging, Tracing, and Catching Quality Drift
You cannot fix what you cannot see. Tracing LLM calls, logging prompts and outputs lawfully, latency and cost dashboards, and alerting on quality regression.
AI Content Moderation in SaaS: Safety Without a Trust-and-Safety Team
Small teams need moderation too. Classifier cascades, LLM-based review, human escalation, and the DSA duties that apply once users generate content.
AI-Driven Onboarding: Personalising the First Five Minutes
The first session decides activation. Using an LLM to tailor onboarding to the user's stated goal, generate starter content, and shorten time-to-value.