Almog Baku

Almog Baku is co-founder and chief executive of Kelet, which diagnoses production AI-agent failures and recommends targeted fixes. He founded GenAI Israel and created the LLM Triangle, a framework for building reliable applications around language models.

Earlier in his career, Baku co-founded Rimoto and contributed to Kubernetes. By 2022, he was developing Raptor, an open-source project for moving Python-based machine-learning research onto production infrastructure. He founded GenAI Israel that year, bringing together engineers, researchers, and founders around practical generative-AI development; he subsequently helped launch a community focused on AI security.

His later open-source projects tackle operational problems specific to language-model applications. openai-streaming simplifies streamed responses, structured outputs, and tool calls. pytest-evals incorporates agent evaluation into familiar pytest and continuous-integration workflows.

How he approaches production AI

• LLM Triangle: His production-engineering framework treats model selection, application architecture, and contextual data as interdependent decisions, guided by explicit standard operating procedures. Engineers should map how human experts actually complete tasks, including implicit intermediate judgments, before translating those steps into software.

• Bounded agent autonomy: Baku favors defining where agents can improvise while keeping surrounding workflows predictable and debuggable. His AI Engineer Summit presentation illustrates the approach with an encyclopedia-writing system that separates research, outlining, and drafting, reserving deterministic tasks for conventional code.

• Focused context and incremental optimization: Relevant examples and carefully selected retrieval often improve reliability more than indiscriminately expanding context. He recommends starting with a capable model, collecting performance data, and optimizing cost, latency, or fine-tuning only after establishing a working baseline.

• Cross-session failure analysis: At Kelet, Baku addresses recurring failures that individual traces can obscure. His analysis of production failure patterns describes identifying systemic problems across interactions, including useful evidence buried beneath irrelevant context, then suggesting targeted changes to prompts or application logic.