My AI stack
Current snapshot · October 14, 2026
This is not a tool ranking. It is the working system I use from voice notes and product discovery to architecture, code, verification, and production. These are technologies I work with.
How the work flows
01 · Voice and idea
Voice and text tools help capture ideas and formulate tasks.
02 · Context
Task management, a text repository, and a knowledge base help prepare development context and rules.
03 · Build
Codex and Claude Code inspect repositories, write code, run tests, and check interfaces. Larger tasks can be split among several agents.
04 · Verification
Code lives in Git. Before release I run builds, tests, and live browser checks. AI features are evaluated with reference examples and LLM error analysis.
Stack layers
Agents
Codex · Claude Code · LLM APIs
Agents are development tools, not substitutes for product decisions. I define architecture, context, and acceptance criteria, then review the result.
Context and tasks
Markdown · text repository · APIs · Git
Organizing requirements, decisions, and project context; integrating sources with controlled access.
Product development
TypeScript · Python · Next.js · React · Astro · HTML/CSS/JS · Figma
Prototypes, web interfaces, internal services, and production products. Visual experiments use p5.js, p5.brush, and WebGL.
AI infrastructure
PostgreSQL · Qdrant · RAG · Temporal · n8n · Payload CMS
PostgreSQL and Payload handle data and content, Qdrant supports semantic search, Temporal runs reliable workflows, and n8n validates multi-step AI scenarios.
Voice and messaging
Whisper · Telegram Bot API · Edge TTS · Swift
Speech recognition and synthesis, voice interfaces, and messaging integrations.
Devices
ESP32-S3 · C++ · Arduino/ESP-IDF · USB · BLE
Voice and HID prototypes: audio capture, USB and Bluetooth keyboards, buttons, encoders, and macros.
Delivery
Docker · CI/CD · Kubernetes · browser testing
Repeatable environments, builds, tests, and live-interface checks. Enterprise solutions are adapted to secure environments and Kubernetes.
Where it is used
SuperSlide →
AI-assisted creation of corporate presentations.
Voice AI Assistant →
Voice and text queries across work sources.
Voice equipment inspection →
Capturing equipment inspection results by voice.
Processing →
Experiments with code-generated graphics and animation.
Principles
- Build a working version and test a real scenario before scaling.
- An agent needs context, boundaries, and acceptance criteria—not just a prompt.
- LLM output needs tests, eval datasets, and human review.
- Secure scenarios require local processing, access control, and observability.