New technologies!

New technology matters when it creates a useful capability, changes an architectural boundary or makes an existing system meaningfully better. I evaluate emerging AI, platform and software technologies by what they actually enable, how they integrate with existing systems and what operational trade-offs they introduce.

Where My Current Technical Focus Is

AI Application Architecture

The interesting layer in modern AI is increasingly the system around the model: provider abstraction, retrieval, context construction, tools, permissions, state, evaluation, observability and runtime control. Models remain important, but application architecture determines how their capability becomes usable.

• Local, hosted and hybrid model runtimes
• Provider and model abstraction
• Capability-aware model routing
• Structured outputs and application APIs
• Evaluation and behavioral validation
• Permission-aware tool execution
• AI clients and controlled runtime environments

Agents, Tools & Protocols

Agentic systems introduce a new application layer in which a model can select tools, maintain state, observe results and continue toward a goal. The important engineering questions concern execution boundaries, permissions, reliability, state management and how capabilities are exposed to the runtime.

Protocols such as MCP and related agent interoperability approaches are relevant because they standardize specific connections between models, runtimes, tools, data and other systems. They should be understood as architectural layers rather than treated as replacements for the complete application.

Retrieval, RAG & Context Engineering

Retrieval-Augmented Generation is evolving from a simple vector-search pattern into a broader retrieval architecture. Embeddings, lexical search, metadata filtering, reranking, Source-of-Truth selection, provenance and context construction are separate concerns that need to work together.

• Embeddings and semantic representation
• Vector and lexical retrieval
• Hybrid search
• Reranking
• Metadata and permission filtering
• Source provenance and evidence
• Context engineering
• Answer validation boundaries

Local, Private & Controlled AI

Local inference is becoming a practical architectural option alongside hosted model APIs. Tools such as Ollama and OpenAI-compatible local runtimes allow applications to choose between local and remote execution according to privacy, latency, capability, cost and operational constraints.

Local AI should not be confused with air-gapped or sovereign AI. Those concepts impose different requirements on infrastructure, dependencies, software supply chains, operations and organizational control.

Structured Knowledge for Humans and AI

Content systems increasingly serve more than a browser page. Structured documents can also support search, retrieval, translation, knowledge graphs, internal linking and AI applications. This makes block identity, metadata, provenance, semantic relationships and multilingual synchronization architectural concerns.

stajic.de uses structured EditorJS content and multilingual publishing as a practical environment for exploring this transition from traditional CMS content toward machine-readable technical knowledge.

Modern Platform Engineering

Modern application stacks combine server-side rendering, typed frontend and backend code, relational data models, modular APIs and production infrastructure. Nuxt, Vue, TypeScript, Node.js, Prisma and PostgreSQL provide a strong current foundation for this class of system when their architectural fit is right.

Desktop AI & Local Runtime Integration

Desktop applications are becoming relevant again as AI workloads need controlled access to local models, files, tools and operating-system capabilities. Aaasaasa AI Client explores this boundary using Electron and TypeScript together with local and remote AI runtimes, provider abstraction and controlled application permissions.

Identity, Permissions & AI Security Boundaries

As AI applications gain access to tools and enterprise data, traditional application security becomes even more important. Authentication, authorization, RBAC, tenant isolation, runtime permissions and auditability must remain explicit boundaries around model behavior.

Production Matters More Than the Demo

A technology demo answers whether something can work. Production architecture asks whether it remains understandable, secure, maintainable and operable when the system has real users, data, failures, migrations and changing requirements.

• Can it be deployed and upgraded predictably?
• Can failures be observed and diagnosed?
• Can data and permissions be controlled?
• Can the technology be replaced if dependencies change?
• Can the team understand and operate it?
• Does its benefit justify its complexity?

Adopt, Evaluate or Experiment

Not every promising technology belongs immediately in production. I separate technologies that are mature enough for normal implementation from those that should be evaluated in a controlled proof of concept and those that are still useful primarily for research and experimentation.

Technology Through Implementation

New technologies are evaluated through implementation wherever possible. stajic.de, Aaasaasa AI Client, SenseFlow, loving.rocks, figure.rocks and bazify.com provide different environments for testing platform architecture, multilingual publishing, AI integration, local runtimes, search, technical SEO and production operations.

The Goal Is Better Systems, Not More Technology

I follow emerging technology closely, but novelty alone is not a design criterion. The technologies worth adopting are those that create useful capability while keeping the surrounding system understandable, controllable and deliverable.