Vladislav Beliaev
RL researcher, founding engineer
7 years building models and AI products end to end.
now: RL post-training research on cross-model reasoning.
bets
AGI is a fusion of maximally different, equally strong models. the lab race is breeding exactly that lineup
labs deliberately keep reasoning verbose. the same labs set the research agenda. that leaves an enormous field untouched
only the boldest ideas are worth a human's time now. the incremental rest, agents will grind out anyway
career
2026
Stealth RL Lab ($10M funded)founding research engineer
competitive cross-model RL, task shaping for hard reasoning, multi-model co-reasoning, and end-to-end RL systems
2024–2026
Fluently (YC W24)founding engineer
post-trained and deployed Qwen3; built the real-time voice agent from scratch at sub-500ms and 50k+ DAU; owned core AI through 0 → $8M ARR
2021–2024
Neiro.aidl engineer → head of ml → head of r&d
built speech models and B2B AI video capabilities, led a 15-person team, and scaled the product to $150k net MRR and 10k+ DAU
2019–2021
JetBrains Researchresearch engineer
led student researchers, published three papers, and built tooling for scientific paper analysis
M.Sc. in ML
B.Sc. in math
papers
Max Out GRPO Signal: Adaptive Trace Prefix Control for Hard Reasoning Problems
accepted to AAAI 2027 · the hardest problems become the strongest RL signal · arxiv
Agon: Competitive Cross-Model RL with Implicit Rival Grading of Reasoning
two models grade each other's reasoning, a step toward latents · arxiv
SpeechNet: Weakly Supervised, End-to-End Speech Recognition at Industrial Scale
trained without human labels, shipped at 12M queries a day · arxiv
End-to-end Deep Object Tracking with Circular Loss Function for Rotated Bounding Box
transformer tracker with a custom loss, beat SoTA on VOT Challenge · arxiv
Transformer-Based Open Domain Biomedical Question Answering at BioASQ8 Challenge
open-domain QA with BERT, before LLMs made it easy · arxiv