AI systems research · agent reliability · post-training · AI safety · mechanistic interpretability
I’m Igor Bogdanov, an AI systems researcher and research engineer. I design compound LLM-agent systems, develop inference-time adaptation methods, and build evaluation infrastructure for long-horizon reliability, grounded in more than 15 years of production software and systems engineering.
A system map with four stages connected top to bottom: Architecture, Adaptation, Evaluation, and Interpretability. Cross-links label Context engineering, Population memory, Time-to-failure, and Reasoning representations. A dashed feedback arrow shows interpretability informing architecture.
Research program
My work examines the full lifecycle of an agentic AI system: how it is structured, how it adapts, how it fails, and what its internal representations reveal.
What should an agent see, and how should decision-making be decomposed?
Structured state and bounded interfaces often deliver more value than simply adding longer deliberation.
Compound Agent DesignCan agents improve through experience without updating model weights?
FORGE turns failed trajectories into reusable memory and uses population-level selection to stabilize prompt-only learning.
FORGEWhen does an agent abandon a goal, and what reasoning precedes the failure?
Time-to-event analysis reveals early impulses, late-stage fatigue, model-specific failure fingerprints, and contradiction inside elaborated rationales.
Multi-Turn ConsistencyDo language models reuse common reasoning structure across languages?
Geometry-invariant sparse autoencoders and causal analysis probe the extent to which reasoning representations are shared and functionally interchangeable.
Cross-Language ReasoningFeatured research
Prompt-only reflection becomes more reliable when useful discoveries are selected and broadcast across a population.
What an agent sees can matter more than how long it deliberates.
Reliability is temporal: models fail at different stages and produce systematically different narratives before doing so.
Geometric alignment across languages does not automatically imply functional interchangeability.
Research to working systems
Each study on this site ran on infrastructure I built: model connectors, harnesses, orchestration, logging, and the evaluation pipeline behind every reported number.
01
02
03
Selected publications
Igor Bogdanov, C.-H. Lung, T. Kunz, J. Gao, A. Taylor, M. Zaman
ACM Conference on AI and Agentic Systems (CAIS), 2026
Igor Bogdanov, C.-H. Lung, T. Kunz, J. Gao, A. Taylor, M. Zaman
ACM Conference on AI and Agentic Systems (CAIS), 2026
Igor Bogdanov, C. Huang
ICML 2026 Workshop on Mechanistic Interpretability
Igor Bogdanov, O. Manakina, C.-H. Lung
ICLR 2026 Workshop on LLM Logical Reasoning
O. Manakina, Igor Bogdanov, C.-H. Lung
NeurIPS 2025 Workshop on Multi-Turn Interactions in Large Language Models
Background
A trajectory, not an employment ledger, newest first: original research on agent systems, formal training, and fifteen years of production software before that.
Education
Contact
I’m interested in research engineering, applied research, agent infrastructure, evaluation, reliability, interpretability, and research-to-production work.