← Back to blog
2026-07-14 · 2 min read
The AI Stack I Use in 2026
AI StackTools
Principles First
Before listing tools, the principles that guide selection:
- Sustainability — Will this tool exist in 2 years? Is it backed by a real company or community?
- Autonomy — Can I self-host, export, or migrate? Am I locked in?
- Cost — Does the pricing model make sense at scale?
- Reliability — Has it survived production use, or is it a demo?
The Stack
LLMs
- Primary reasoning: Claude (Anthropic) — best for complex analysis and code generation
- Fast tasks: GPT-4o-mini for cheap, fast summarization and classification
- Open-source fallback: Llama variants via cloud inference for cost-sensitive workloads
Orchestration
- Agent framework: Custom — I've tried most of the popular ones and found that a lightweight custom framework gives more control and fewer surprises
- Workflow automation: n8n for visual pipelines, custom code for everything else
Data Layer
- Vector storage: SQLite with vector extensions — simple, portable, no external dependency
- Document storage: File system + Git — Markdown files are the most future-proof format
- Structured data: PostgreSQL for anything that needs real queries
Frontend
- Web: Next.js + Tailwind — battle-tested, great DX, excellent SEO
- Content: Markdown + remark for blogs, structured JSON for everything else
Deployment
- Hosting: Netlify for static, Vercel for dynamic — both have excellent free tiers
- DNS: Cloudflare — free, fast, reliable
- Monitoring: Uptime checks + log-based alerting
Communication
- Delivery: Telegram for real-time, email for formal
- Voice: TTS for briefings when hands-free is needed
What I Dropped
- LangChain — too many abstractions, too much magic, too many breaking changes
- Pinecone — expensive, vendor lock-in, SQLite works fine for my scale
- Mid-tier CMS platforms — Markdown in Git is simpler and more portable
What I'm Watching
- Local inference getting cheaper — may shift more workloads on-device
- Agent-to-agent protocols — standardized communication between agents
- Structured output improvements — making LLMs more reliable for data tasks
The stack isn't exciting. That's the point. It works, it's sustainable, and it doesn't change every week.