SPOT — AI-Powered Basketball Video Analytics
Computer Vision & MLOps
May 2026 – Present
- Built an end-to-end AI system that transforms full-length basketball footage into structured scoring events, player-level analytics, shot maps, and automatically generated highlight clips, owning the complete product from computer vision research and model development to distributed inference infrastructure and the user-facing web application.
- Developed the complete perception stack around five task-specific models, covering ball and rim detection, court understanding, action recognition, and jersey-number OCR, with custom data collection and annotation, fine-tuning, evaluation, and task-specific post-processing rather than relying solely on off-the-shelf inference.
- Designed custom computer-vision algorithms beyond learned models, including court homography, appearance-based team separation, roster-aware jersey identification, and a physics-constrained global trajectory solver using dynamic programming to recover ball trajectories and suppress persistent false positives.
- Built a distributed serverless GPU inference architecture, decomposing full-match processing into parallel workloads across specialized GPU containers and combining model inference with CPU-side processing, storage, orchestration, and asynchronous job execution.
- Iteratively optimized the system against manually labelled ground truth, replacing assumption-based performance estimates with measured evaluation and improving scoring detection while substantially reducing full-match inference cost through hardware profiling, workload placement, and candidate-ranking strategies.
- Shipped the complete production product as its Founding AI Engineer, combining the Python vision infrastructure with a Next.js/React application, authentication, row-level data security, cloud object storage, deployment infrastructure, and a workflow that turns raw match uploads into browsable analytics and clips.
PythonPyTorchRF-DETRPARSeqSAM2OpenCVModalCUDANext.jsReactTypeScriptSupabaseCloudflare R2Railway
EComGen — AI Product Content Platform for E-Commerce
Founder · Full-Stack AI / Generative AI
2025–2026
- Built an end-to-end Generative AI SaaS for e-commerce sellers, enabling merchants to connect Shopify or Amazon stores—or work independently—and transform existing product assets into AI-generated product imagery, videos, and multilingual SEO content from a unified platform.
- Designed the complete AI generation workflow around Google Vertex AI, integrating multimodal Gemini models for image and text generation and Veo for product video generation, alongside custom prompting, brand-aware visual generation, watermarking, product-feature overlays, and multilingual content generation.
- Integrated the product directly with commerce platforms, implementing Shopify OAuth and Amazon SP-API workflows for catalog synchronization and enabling generated assets to be published back to product listings without leaving the platform.
- Built the production SaaS infrastructure end-to-end using React, TypeScript, Express, PostgreSQL/Drizzle, Firebase Authentication, Google Cloud Storage, and Railway, including background job processing, concurrency controls, transactional credit accounting, administration tooling, and Stripe-based payments.
- Developed a programmatic SEO system generating 5,000+ server-rendered landing pages, spanning product categories, commerce platforms, industries, use cases, guides, and glossary content across multiple languages, supported by dedicated sitemap infrastructure and structured metadata.
- Took the project from empty repository to a deployed, publicly accessible product, owning product design, AI integration, backend and frontend development, cloud infrastructure, database architecture, marketplace integrations, payments, SEO, and deployment as a solo developer.
TypeScriptReactExpressPostgreSQLDrizzle ORMVertex AIGeminiVeoFirebaseGoogle Cloud StorageShopify APIAmazon SP-APIStripeRailwayTailwind CSS
Deep Learning for Dementia — Hackathon '23
2nd Place · TechCrew 24-Hour Hackathon
Aug 2023 – Aug 2023· 1 mo
- Developed a deep-learning image classification prototype for dementia classification from brain MRI scans, working with a four-class, approximately 6,400-image dataset under a 24-hour hackathon constraint.
- Iterated on preprocessing, architecture design, and model configuration to reach 91% classification accuracy, then wrapped the resulting model in a Gradio-based interactive application for direct image upload and inference.
- Placed 2nd in the hackathon, demonstrating the complete prototype from model development through an accessible inference interface within the competition window.
PythonTensorFlowKerasComputer VisionGradio
Real-Time Object Recognition — TÜBİTAK Project
Computer Vision & Deep Learning
Nov 2022 – May 2023· 7 mos
- Developed and evaluated a real-time object detection pipeline across a custom 100-class image dataset, working through dataset curation, annotation, preprocessing, training, evaluation, and real-world inference.
- Experimented with YOLOv5 and multiple CNN architectures, building on earlier classification work with VGG, ResNet, EfficientNet, and Inception-family models to transition from image-level recognition toward real-time object localization and classification.
- Applied image preprocessing and dataset-quality techniques including histogram equalization and systematic annotation, and evaluated multiple YOLO model configurations for the accuracy/latency trade-off required by real-time inference.
- Integrated the resulting models with PyTorch and OpenCV for live inference, achieving substantially stronger real-world recognition performance than earlier iterations of the project.
PythonPyTorchYOLOv5OpenCVCNNsComputer VisionData Annotation
Image Recognition & Deep Learning — TÜBİTAK Project
Computer Vision Research
May 2022 – Nov 2022· 7 mos
- Worked on the earlier image-classification stage of a 100-class computer-vision project, contributing across dataset curation, manual annotation, preprocessing, model training, and comparative experimentation.
- Evaluated multiple deep-learning architectures including AlexNet, VGG16/19, EfficientNet, Inception/GoogLeNet, and ResNet variants, comparing their behavior and adapting training configurations to the domain.
- Developed preprocessing workflows including histogram equalization and glare reduction, improving dataset consistency before training and establishing the foundation later used for the project's real-time YOLO-based recognition work.
PythonTensorFlowKerasOpenCVCNNsImage Processing
