Work
Making LLMs reliable and cheap to run in production.
Public sources, organized into notes on harness engineering, LLM cost, and agentic systems — the same problem space I work in day to day.
About_me
I'm Bartłomiej Krupa, Head of Agentic Engineering at Brand24, with 10+ years in IT. Day to day I optimize LLMs, roll out agentic engineering across engineering teams, and research how these models actually behave.
What_I_work_on
LLM & cost optimization
Profiling token and dollar spend on live LLM and agent features, then cutting it — caching, model routing, prompt and context compression — without losing quality, measured against evals.
AI-assisted coding adoption
Rolling out agentic engineering across teams: tooling, guardrails, and workflows that coordinate agents without lowering the quality bar.
Agentic harness & LLM behavior
Validated, multi-layered agent systems with real production access — context pipelines, feedback loops, fallbacks, and monitoring — informed by hands-on research into how LLMs behave.
How_I_work
Architectures over models
Never trust raw output. Validation, feedback loops, and fallbacks — LLMs are unreliable; architectures are not.
Cost-aware by default
Token and dollar budgets, caching, and model routing — measured against evals, not vibes.
Proof
First public case studies are in progress. References from recent work are available on request — email me with your workflow and I will show you exactly how I would approach it.
Working on the same problems?
If you're cutting LLM cost, building agentic systems, or bringing AI-assisted coding into your team, I'm happy to compare notes.
Email me