AP
ALEŠ POGAČAR
SOCIAL · TECHNICAL · IMPACT DRIVEN
← All experience views
Agentic AI & Automation

Multi-agent delivery systems — and the automation before them.

The newest chapter: using Claude Code and Cursor as coordinated agents to design, build and ship production software. Underneath it is a decade-long habit of automating anything repetitive — including the IoT support systems that built the instincts behind it.

Lead Architect — Agentic AI Delivery & LLM Product
Private Product (GCP) — Remote
2025 – 2026
  • Two AI agents + human: Claude Code plans/reviews; Cursor builds/deploys; human relays and approves.
  • Git-only state: specs, decision log, handoff threads, two-gate PRs (plan review, then code review).
  • Production LLM app on Cloud Run — RAG, safety gates, escalations, admin portal.
  • Channels: WhatsApp/Twilio; 17/17 end-to-end checks passed; 15+ defects caught before launch.
  • Automation stack: n8n workflows + Python services + Cloud Build CI/CD.
Agentic AIClaude CodeCursorLLM / RAGn8nGCPPython
AI guest assistant — RAG SaaS design →
Platform Lead — Multi-Agent Dev Platform (Claude Code & Cursor)
Private Research Codebase — Remote
2024 – Present
  • Developer → independent reviewer → instructor loop with JSON-schema-validated handoffs; the reviewer re-runs all evidence itself, refute-by-default.
  • Runs unattended: durable resume ledgers (one committed row per task), parked tasks that never block the batch, and singleton resources (Sierra Chart) serialized automatically.
  • Cost discipline: per-task model tiers (light / standard / critical) — "economize on generation, never on the gate."
  • Proven at scale: this system built and operates a ~334k-line quant research platform and its ~90k-line C++ execution layer, with peaks of 50+ independently reviewed commits per day.
  • Evolved from an earlier Cursor-based setup (infrastructure agent owning shared libs, strategy agents in isolated folders, git as memory between sessions); the same protocol was reused for the production LLM product above.
Claude CodeMulti-Agent WorkflowAgentic AICursorPythonGit Handoffs
Multi-agent engineering system → The quant pipeline it built →
Systems Engineer — IoT, LLM Support Automation (Python)
Koren Sports d.o.o. — Slovenia / Austria / Multiple Venues
2021 – Present
  • Built a camera system for Planica (world's largest ski jump): GoPro Max 360° cameras record zipline descents, auto-triggered via Raspberry Pi and Bluetooth LE.
  • GoPro Hero 12+ automatic photo capture at adventure parks with the same central-management approach.
  • Central management over MikroTik routers and WireGuard VPN tunnels: remote monitoring, SSH, file pull and watchdogs.
  • Automated customer-support email responses using LLM APIs — the earliest production use of LLMs in this body of work.
PythonRaspberry PiBluetooth LEWireGuardMikroTikLLM APIs
Adventure park camera system →