Projects

Two kinds of work: self-directed machine learning projects with public code and results, and software I have built professionally.

Research & learning

Projects I use to build a foundation for machine learning research. Every result shown comes from a run I actually made.

imbalanced-benchmarks

Imbalanced Classification Benchmarks

Comparison of cost-sensitive weighting, SMOTE and random undersampling against untreated baselines on three medical datasets, using 5-fold stratified cross-validation scored with PR-AUC instead of accuracy.

Committed result: balancing methods mostly improved recall rather than ranking, and made little difference when imbalance was mild.

scikit-learn imbalanced-learn cross-validation
vision-baselines

Vision Baselines

Reproducible PyTorch training harness for CIFAR-10 with LeNet and a CIFAR-stem ResNet-18: deterministic seeding, cosine learning-rate schedule, per-epoch metrics logged to CSV.

LeNet reaches 80.25% test accuracy in 30 epochs; the ResNet-18 run is in progress.

PyTorch CIFAR-10 reproducibility
semantic-image-search

Semantic Image Search

Text-to-image search over local photo collections using pretrained CLIP embeddings, with an incremental embedding index and a command-line interface for natural-language queries.

Working retrieval tool; index updates without re-embedding unchanged images.

PyTorch CLIP embeddings
software-foundations

Software Foundations

Worked exercises from Software Foundations Volume 1 in the Coq proof assistant: functional programming and proofs by induction, checked by the proof checker.

Ongoing; Basics chapter complete.

Coq formal proofs

Professional work

Products and platforms I have worked on as a software engineer.

Screenshot of Navian Platform Navian Platform Full-stack engineer Real estate analytics platform with market intelligence reports, subscription billing, and investment tools for the Swedish market. TypeScript Node.js React PostgreSQL Python Screenshot of OceanDrive OceanDrive Full-stack engineer Decentralized storage platform for peer-to-peer file distribution across networked nodes. React Next.js Node.js PostgreSQL Screenshot of World Art DEXPO World Art DEXPO Frontend engineer Digital art marketplace built with React and NestJS, with wallet-based authentication and on-chain asset ownership. React NestJS Web3 Screenshot of ResumeForge ResumeForge Solo developer Resume builder with ATS-friendly templates, real-time preview, and payment integration. Built and operated as a solo side project. Next.js TypeScript Stripe
MirzoDev

© 2026 Islomjon Mirzakbarov

Email GitHub