Skip to content
View mbrcic's full-sized avatar

Block or report mbrcic

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
mbrcic/README.md

Mario Brcic

AI alignment: stating what AI must do, and checking it by machine.

Associate Professor at FER (University of Zagreb). The work has four parts: stating exactly what AI must do, building the tools that check it (formal verification, interpretability of language models), finding the limits no alignment method can get past, and asking who sets AI's goals.

Current projects

  • AI Safety Formalization Atlas — an open Lean 4 library of machine-checked AI-safety mathematics, with every result's source and scope on record. Documentation.
  • VALOR — FER research project on value alignment and interpretability of large language models. First output: llm-political-steerability, the reproducibility package for Auditing Alignment Controllability in LLMs via Political Axes (AIES 2026).
  • Modern Algorithm Design — a textbook on designing algorithms from reusable primitives, and on stating precisely what an algorithm must guarantee. Its code, with each structure in Python, Rust and Lean, is in algodesign.

Selected upstream contributions

Notable earlier work

  • Explainable optimization (2018) — tracking-predictive-gantt: a visual-explainability method for stochastic scheduling (ECDF uncertainty and counterfactual overlays). The failure modes it exposed led to a follow-up scheduler with gains of up to ~770% over the prior state of the art. DOI

The pinned repositories are the packaged, reproducible work. The others range from active research to teaching material, at varying polish. Questions and collaboration are welcome.

Pinned Loading

  1. ai-safety-formalization-atlas ai-safety-formalization-atlas Public

    The open workbench for AI safety, made formal. Turn safety questions into machine-checked Lean proofs — a shared launchpad where researchers and AI agents build provable safety together.

    Lean 14 4

  2. llm-political-steerability llm-political-steerability Public

    Reproducibility package for 'Auditing Alignment Controllability in LLMs via Political Axes' (AIES 2026): raw data + collection/analysis pipeline.

    Python 3

  3. algodesign algodesign Public

    The executable companion to Modern Algorithm Design. Each algorithm, three ways: Python to read and run, Rust to engineer, Lean 4 to state and prove what it promises.

    Lean 5 1

  4. Deepseek-OCR-vllm-docker Deepseek-OCR-vllm-docker Public

    Self-contained, high-performance DeepSeek-OCR service powered by vLLM and FastAPI. Zero-friction Docker setup for images, PDFs, and batch processing.

    Python

  5. tracking-predictive-gantt tracking-predictive-gantt Public

    Early XAI-for-optimization: visual explainability for stochastic scheduling (ECDF + counterfactual Gantt, 2018). Basis for an EJOR method beating SOTA by up to ~770%.

    Python