AI Research & Engineering
Spectral reasoning, tensor compression, multi-modal benchmarks, and self-improving agents
Active work across reasoning, benchmarks, optimization, and meta-governance.
Research pipeline protocol — idea to publication. Telegram bot (19 commands), multi-agent orchestration, CI/CD, 7 pipeline stages.
Latent energy-based GRPO reasoning with spectral rewards on LFM2.5. Active research.
Cross-modal dense context reasoning benchmark. MERA integration for Russian LLM evaluation.
Bayesian black-box optimization framework. General-purpose optimizer for hyperparameter search.
Differentiable tensor compression via Tucker/HOSVD in JAX. GPU-accelerated scientific data compression with learned factor matrices.
Interactive Hebrew textbook for Russian speakers. 19 lessons: alphabet, grammar, vocabulary, quizzes, flashcards. Aleph level, zero dependencies.