Things I've built and tested.

A reproducible, leakage-safe evaluation pipeline for U-Net-family models on multimodal brain-tumor MRI, built for the TÜBİTAK 2209-A project.

Problem

Segmentation results in medical imaging are hard to trust when train/test splits leak between patients, evaluation protocols shift between papers, or preprocessing choices go undocumented.

Approach

Built the evaluation around a fixed patient-level split, an audited dataset, and a written analysis plan frozen before the held-out test set was touched — then ran one guarded internal evaluation against it.

Engineering

Implemented and compared standard U-Net, BU-Net and U-Net+RES under a bounded, matched 2D training protocol; ran multi-seed confirmation for stability; tracked measured compute per model; produced every figure and table from tracked artifacts; verified the whole thing from a clean clone; assembled the manuscript package with an internal review round.

Stack

TensorFlow, Keras, Python, pytest.

Result

On the 74-patient held-out subset, seed-aggregated mean regional Dice was 0.736 for standard U-Net, 0.752 for BU-Net and 0.756 for residual U-Net. Both variants beat the baseline under Holm-corrected paired bootstrap tests; the gap between the two finalists was small and borderline after correction. Residual U-Net reached its score with fewer parameters, fewer per-slice MACs and lower peak VRAM. Internal single-dataset evidence — no external validation, and no clinical claim.

A role-based education and business-operations platform connecting sales, learning delivery, assessment and reporting in one product.

Problem

Running exam-prep education — lessons, mock exams, CRM and operational reporting — across disconnected tools creates friction between teaching, sales and operations.

Approach

One platform, role-based access: lesson delivery and progress tracking, a mock-exam club, and CRM/ops reporting behind a single account system.

Engineering

CI and Lighthouse performance checks run on every change via GitHub Actions; the app ships as a container image to GHCR.

Stack

Next.js, TypeScript, Prisma, PostgreSQL, Playwright, Docker.

Result

Live in production at onlinedershanem.com.

The subscription and account layer behind OneArticle — one human-reviewed article briefing delivered every weekday.

Problem

Ship a small, sustainable subscription product without hand-rolling billing, verified sign-up and transactional email from scratch.

Approach

A single verified product path: email-code sign-up, reader preferences, checkout, then scheduled delivery to eligible subscribers only.

Engineering

Checkout, subscriptions and webhooks via Polar; verification, editorial and alert email via Resend with RFC 8058 one-click unsubscribe and bounce/complaint-based suppression; daily send via a Vercel cron job; error monitoring with Sentry.

Stack

Next.js, TypeScript, PostgreSQL, Prisma, Polar, Resend.

Result

Live production subscription product with real billing and delivery infrastructure.

A daily-games sibling to OneRead — OneSudoku, OneDna and OneWord under one shared product shell.

Problem

Extend the same product family into daily casual games without fragmenting the brand or rebuilding the shell from scratch.

Approach

Reused OneRead's design system — typography, layout, lockup, footer — so both products read as one company, while shipping three independent daily puzzles.

Engineering

Email-code verification and subscription billing shared across games, plus an explicit no-payment test path; consistent page padding and lockup placement verified by tests.

Stack

Next.js, TypeScript, Prisma, Polar.

Result

Live production product, sibling to OneRead.

Symbolic Python verification of quantum group structures — representations, Hopf axioms and Yang–Baxter relations.

Problem

Verify, by computation rather than by hand, the algebraic claims behind a piece of mathematical research on quantum group structures.

Approach

A tightly scoped symbolic toolkit for U_q(sl₂) and GL_q(2|1) — not a general theorem prover, but a way to pin the paper's finite-dimensional representations and relations to reproducible tests.

Engineering

Explicit E, F, K, K⁻¹ matrices for finite-dimensional representations; representation-level zero-residual checks of the defining relations; Hopf-structure axiom checks (comultiplication, counit, antipode); a 9×9 R-matrix and graded Yang–Baxter verification for GL_q(2|1).

Stack

Python, SymPy, pytest.

Result

A reproducible, test-verified companion codebase to the underlying mathematical research.