Production AI Agent Tech Stack

A Reality Check: Beyond Calling GPT-4 in a For-Loop

Based on building autonomous schedulers, multi-agent research assistants, and several €100Ms value creation across 10+ companies and sectors

View All Frameworks

Critical Reality Check

LLMs are not the solution to all agentic use cases. Custom ML models and reinforcement learning systems remain the workhorses of production AI. LLMs expand the economic frontier of what's worth building—tasks that previously required six months of data collection can now be prototyped in a weekend.

The art is knowing which tool solves which problem—and when to combine them.

Four Guiding Principles

1Research-driven, but pragmatic
2Best tool for the job (framework agnostic)
3Observability from day one
4Production-ready beats clever

What I'm Watching

  • Inference-time compute scaling (o1-style)
  • Compound AI systems
  • Guardrails AI for safety boundaries
  • DPO and RLAIF for alignment
  • Mamba and State Space Models

Tech Stack Philosophy: Research-driven execution with production-ready patterns

Built from experience spanning autonomous systems, multi-agent architectures, and enterprise AI at scale