CV

Pavel Golikov

AI Researcher & Agent Engineer

paulgolikov@gmail.com
Toronto, Ontario, CA

Summary

AI Researcher and agent engineer specializing in robust machine reasoning, retrieval-augmented generation (RAG), LLM evaluation, and alignment. Builds grounded agentic workflows, vector-retrieval systems, and dataset generators for context-management evaluation in LLM agents. Former Military Intelligence Operator who held a Top Secret clearance, bringing a threat-modeling mindset to AI security and model capability evaluation.

Education

  • PhD in Computer Science (leave of absence beginning September 2026)
    Present
    University of Toronto
  • Master of Science in Computer Science
    2022
    University of Toronto
  • Bachelor of Science in Mathematics and Philosophy (Formal Logic)
    2011
    University of Toronto

Work Experience

  • Graduate Researcher (PhD, on leave) — ML Systems & Agent Evaluation
    2022 - Present
    University of Toronto, EcoSystem Research Group
    • Conducted research spanning ML systems and efficient computation, with later work focused on robust machine reasoning, agent evaluation, context management, and AI alignment, in the EcoSystem Research Group and as a member of the Vector Institute.
    • Developed open-source evaluation frameworks, agentic RAG systems, and mechanistic-interpretability analyses, resulting in first-author research on context management and reasoning robustness.
  • Graduate Researcher (MSc) — Distributed Systems
    2020 - 2022
    University of Toronto, EcoSystem Research Group
    • Engineered a flexible IoT distributed data-streaming framework from scratch, designed to automatically partition computational streaming queries between edge devices and cloud instances.
    • Built the full software stack: programmed Arduino/C++ sensors for real-time biological data collection (EMG/ECG), developed custom socket networking protocols, and deployed cloud infrastructure using AWS and Apache Flink.
  • Intelligence Operator
    2013 - 2018
    Canadian Armed Forces
    • Formerly held a Top Secret security clearance while conducting rigorous analysis of classified information streams to produce actionable intelligence reports for command elements.
    • Developed a strong adversarial threat-modeling mindset, emphasizing operational security, rigorous data validation, and the identification of logical vulnerabilities in complex, multi-agent scenarios.
  • Mathematics Teacher
    2012 – 2013 & 2018 – 2019 -
    Blyth Academy
    • Taught foundational mathematics to students in Grades 10, 11, and 12, developing the ability to distill and communicate complex quantitative concepts.

Skills

Languages

  • Python
  • C++
  • Java
  • SQL
  • LaTeX

AI & Agents

  • PyTorch
  • Reinforcement Learning (RL)
  • RLVR
  • LangGraph
  • LangChain
  • RAG
  • Hugging Face
  • vLLM
  • Transformers
  • Gemini/OpenAI/Anthropic APIs

RAG & Evals

  • Vector storage
  • embeddings
  • BM25
  • hybrid retrieval
  • cross-encoder reranking
  • MRR/nDCG/recall
  • groundedness
  • validation

Systems

  • Linux
  • SQLite
  • GitHub Actions CI
  • AWS
  • Apache Flink
  • Distributed GPU Clusters

Publications

  • ArbiGraph: Arbitrarily Scalable Verifiable Task Graphs for Evaluating Context Management
    2026
    arXiv
    A benchmark generator for exact evaluation of state management across controllable, executable task graphs. The public implementation also includes reinforcement learning with verifiable rewards (RLVR) environments with dataset-backed and on-demand episodes, hidden verifier state, exact binary rewards, and per-node diagnostics.
  • Robust Reasoning Benchmark
    2026
    arXiv
    An adversarial evaluation of structural fragility in LLM reasoning.
  • Fusing Adds and Shifts for Efficient Dot Products
    2026
    IEEE Computer Architecture Letters
    An algorithmic and hardware optimization for efficient dot-product computation.
  • GPUPool: A Holistic Approach to Fine-Grained GPU Sharing in the Cloud
    2022
    PACT
    A system for fine-grained GPU sharing in cloud environments.
  • Habitat: A Runtime-Based Computational Performance Predictor for Deep Neural Network Training
    2021
    USENIX ATC
    A runtime-based performance predictor for DNN training workloads.

Portfolio

  • ArbiGraph
    2026
    Agent evaluation
    Open-source benchmark and dataset generator for evaluating context management in tool-assisted LLM agents across arbitrarily scalable, long-horizon task graphs. It includes dataset-backed and on-demand reinforcement learning with verifiable rewards (RLVR) environments with hidden verifier state, exact binary rewards, and per-node diagnostics.
  • arXiv Research Agent
    2026
    Agentic ai / rag / llm evals
    Evaluation-first LangGraph literature-review agent with persistent Chroma vector storage, dense/BM25/hybrid/cross-encoder retrieval, citation grounding, resumable SQLite checkpoints, hand-labeled eval datasets, and 163 offline CI tests.