Zhyrgalbek
Kalykov
one project at a time.
Projects
Featured
Vixa AI: Real-Time Hazard Detection
A real-time vision backend for assistive navigation on smart glasses. Fuses YOLOv11 object detection, Depth Anything V2 monocular depth estimation, optical-flow heading, and IoU tracking into per-hazard time-to-collision and danger scoring with directional audio alerts. Includes WebSocket live video streaming, an on-demand multimodal reasoning agent, and a clip-based evaluation harness with automated scoring.
SpeedList AI
Production SaaS that automates e-commerce listing uploads using multimodal AI. Generates trend-optimized SEO metadata at scale with JSON-schema-enforced LLM outputs and Google Search Grounding. Built with async in-memory image processing, real-time Server-Sent Events streaming for AI reasoning logs, and Stripe subscription billing. Cuts 90% of manual work and saves customers 2,000+ hours per month.
Canvas Reminder Bot
A Telegram bot that helps students track their Canvas LMS assignments with smart deadline reminders. Features enterprise-grade security using Google Cloud KMS for token encryption, Firestore for storage, and automated notifications at 24h, 12h, 6h, and 1h before deadlines. Deployed on Google Cloud Run.
Machine Learning
WaterPump Intelligence - ML on Tanzanian Water Pumps
Team ML project on the DrivenData 'Pump it Up' competition: predicting whether 59,400 Tanzanian water pumps are functional, need repair, or broken. Built the full pipeline: regional GPS and population imputation, MinMax normalization, Random Forest vs Decision Tree with multi-class ROC/AUC, and KMeans + PCA clustering, targeting 83%+ accuracy.
Generative AI
Sonungo: AI Job Autofill & Tracking
An AI Chrome extension that autofills complex ATS application forms using a producer-consumer DOM pipeline, semantically mapping unstructured form fields to parsed resume data. Automates end-to-end application tracking by parsing Gmail inbox activity. Cuts job application time by 99%.
Screen-Aware Assistant - AI Overlay for Your Desktop
A screen- and audio-aware AI assistant: an invisible always-on-top desktop overlay (hidden from screen shares) plus a local browser app with real-time answers, live meeting notes, and history. Powered by Claude.
Pretalk: Multi-Modal AI Voice Assistant
A multi-modal AI voice assistant that enhances voice models with speaker diarization, background noise detection, speech-to-text via OpenAI Whisper, and text-to-speech synthesis. Includes RAG-based retrieval for intelligent responses and context-aware embeddings.
UNIVERSITY AI - AskNAU
Developed an AI Q&A model using NLP and machine learning with Python, helping students at North American University get real-time answers to queries. Built the front-end using React and styled with Tailwind CSS for seamless user interaction. Optimized the model with NAU-specific data to improve response accuracy and relevance.
SuperPod - AI-Powered Podcast Analyzer
Built an AI podcast analyzer at the Llama 4 Hackathon in Seattle with a team. SuperPod uses the Llama-4-Maverick-17B-128E-Instruct-FP8 model to transcribe audio, perform semantic search with ChromaDB, and let users jump to specific parts of podcasts using natural language.
Productivity Tools
CouchPilot - Your Mac from the Couch
Run your Mac from the couch: a real terminal plus full screen, mouse, and keyboard control served to your phone's browser over the local network. QR-code pairing, HTTPS on LAN, no cloud, no accounts.
Hackathons
NexHacks @ Carnegie Mellon University
Built Polywatcher, a data analyzer for Polymarket. Selected as a Top 10 finalist! Team: Asel Torogulova, Spencer Chin, and Sergei Ivanov.
Llama 4 Hackathon in Seattle
Built a tool that integrates AI into podcasts for real-time analysis and insights. Worked with an amazing team: Nand Dave, Vishal Shah, and Saptak Sen. Big thanks to Meta for Developers and Cerebral Valley!
Work Experience
Data Engineer Intern @AiKYNETIX
Feb 2024 – Dec 2024 · Houston, Texas
Built a computer vision pipeline for automated movement detection using PyTorch, OpenCV, ViTPose, and MediaPipe, increasing model accuracy by 70% through optimized preprocessing and feature engineering. Automated ingestion, cleaning, and annotation of large-scale video datasets, cutting manual processing by 90%. Developed FastAPI backends integrating LLM and NLP workflows to deliver personalized insights. Debugged video-based model failures through systematic edge-case analysis, and collaborated with engineers and data scientists to ship production ML features.
Education
North American University
Bachelor of Science in Computer Science
Studying core topics in Software Engineering, Artificial Intelligence, and Data Structures. Actively involved in AI projects and hackathons.