The tools I reach for, and why. The list is a habit, not a limit: if you have your own stack, I adapt.
Languages
TypeScript, Python, SQL, Rust
AI and language models
OpenAI, Anthropic Claude, Ollama, Qwen3, bge-m3, hybrid search (vectors + BM25), rerankers, knowledge graph, MCP, n8n, Make
Web and apps
Next.js, React, NestJS, Node.js, Tailwind, shadcn/ui, Expo, Three.js
Data and infrastructure
PostgreSQL, pgvector, Prisma, Redis, BullMQ, Turborepo, Vercel
AI coding tools
Claude Code, OpenAI Codex, ChatGPT
Choices I make on purpose
Code computes, the model reads
Amounts, dates, limits and deadlines are computed by code. The model reads, sorts and suggests. That's why ClaimPilot didn't invent a single number in 54 letters.
Test set first, tuning second
Without a set of questions with correct answers you can't tell whether a change helped. I build it as the first thing in a project.
A local model only when it makes sense
I pick a local model when data must not leave the company or the volume makes it pay off. Otherwise an EU-region API is cheaper and stronger.
pgvector before a separate vector database
If you already run PostgreSQL, pgvector means one less system to manage and back up.
A graph only where it helps
In my measurements a knowledge graph made retrieval worse on ordinary questions (MRR 76.1 → 54.8). So it only runs for questions about relations between documents.
A person approves irreversible steps
Sending an e-mail, a payment or a change in a system is approved by a person until the numbers show the agent can handle it alone.