About Me

I build deployed AI systems and conduct independent reinforcement learning research. My work spans computer vision for retail loss prevention (Sentra Retail OS, co-founded), edge AI on hybrid Radxa/Hailo + Vast.ai infrastructure, and world-model failure modes in model-based RL (DreamerV3 research, targeting RLC 2026). I care about edge AI, real-world deployment, and democratizing AI access in Central Asia.

Service & Professional Activities

  • Founder & Instructor — AI Course: Designed ML curriculum for high-school students; multiple participants qualified for the IOI/IOAI national teams; one cohort member won the Meta LLaMA Accelerator.
  • Organizer & Chief Judge — World Space Olympiad: Co-organized international competition with 200+ teams and a 1,000,000 KZT prize pool.
  • Mentor & Coach — Aqtobe Robotics: Mentored 4+ FIRST teams to international competition stages, including teams that reached FTC Houston Finals.
  • Program Committee Member — YFS (Youth Founders Summit): Entrepreneurship summit with a 6-month accelerator; ~1% acceptance rate.
  • Samsung Innovation Campus (SIC): Advanced AI and neural network research training.
[Note]: Open to research collaborations in RL, world models, and edge computer vision.

Research

  • "When the World Model Lies" — DreamerV3 Failure Modes: Independent research on world-model hallucination and value-function divergence in model-based RL. Findings include imagined-to-real reward ratios reaching ~50× on AntMaze at 500k steps and KL divergence collapse as a leading indicator (r=−0.91, ~50k step lag). Published on Zenodo with reproducible HuggingFace Space and GitHub repository. Targeting arXiv submission and RLC 2026.
  • Polygence Research Fellow (2025): Selected with partial-tuition scholarship (~40% acceptance); conducted mentored ML research project.
  • JOAI 2026 — Neural Decoding: Placed 27/98 in international competition predicting mouse lever position from calcium imaging data.

Ventures

  • Sentra Retail OS — Co-founder & ML/CV Engineer. Hybrid edge/cloud computer vision platform for retail loss prevention, deployed in Astana retail locations. 8-model pipeline (YOLOv11-Pose, AdaFace, SCRFD, MobileCLIP, SigLIP, OSNet ReID, ShopliftingNet 3D CNN, MediaPipe) on Radxa Cubie A7A edge hardware with Vast.ai cloud inference. FastAPI/PostgreSQL/Qdrant backend on Railway, WhatsApp Business API for real-time alerts. Secured 1.2M KZT investment from Almaty Akimat, partnership with ALEMAI, patent granted.

Projects

  • NeuroForge: First UTM drone management system in Kazakhstan.
  • BCI/EEG Wheelchair Control: Real-time EEG pipeline using AD620 instrumentation amplifier; CSP/EEGNet classification with SSVEP and Motor Imagery paradigms. Dashboard built in HTML/vanilla JS/Chart.js.
  • CubeSat Flight Software: ESP32 WROOM + ESP32-CAM platform with LoRa 433MHz (SX1278) telemetry pipeline. Organized World Space Olympiad CubeSat tournament under ALEM Championship.
  • KazPunct (XLM-RoBERTa + CRF): Kazakh punctuation restoration pipeline for hackathon; reached 0.73 F1 on out-of-distribution test data.
  • KazOAI 2026 (Qwen3 LoRA): LoRA fine-tuning pipeline for Qwen3-0.6B/0.8B on structured output generation (JSON/YAML/XML/CSV/TOML).
  • ChronoSort (FLL Project): AI archaeology assistant using SigLIP for artifact classification on Raspberry Pi 5.
  • Bionic Prosthesis: Brainwave-controlled (EEG) artificial limb.
  • Underwater Drone: AI vision for human identification underwater.
  • Smoke Detection: YOLOv5 pre-trained model on Raspberry Pi/ESP32.
  • Intellectual Helmet: VR system and smart sensor for pilots.
  • Forest Fire Monitoring: Ensemble learning pipeline for early fire detection.

Honors and Awards

  • Kazakhstan Republican AI Olympiad — Bronze; National AI Team Selection; 15/725 (Top 2.1%).
  • INFOMATRIX-ASIA 2026 — Silver, AI Programming; 1,135 teams from 21 countries.
  • Decentrathon 5.0 — 4th Place out of 2,300 teams (Top 0.17%).
  • Samsung Solve for Tomorrow — Semifinalist; 38/885 teams selected (Top 4.3%).
  • KAZROCKETS Satellite Design Tournament — 4th Place out of 165 teams (Top 2.4%).

Publications