Production Software Portfolio

Real Projects, Built and Shipped

Each case study names the client and covers the problem, what was built, the technology and the current status, with screenshots or video where the work can be shown.

Manufacturing
ERP
Next.js
Supabase
PostgreSQL
LOGO ERP
Enterprise

SPSTARPLAST Manufacturing ERP

The challenge: An Istanbul manufacturer of insulation anchors needed one system for sales, production, inventory, quality, shipping and purchasing.

What was built: An ERP on Next.js 15 and Supabase: 16 modules and 50+ screens, integration with the company's LOGO ERP, a company-wide activity feed and per-user access control.

Where it stands: In production. Handed over in September 2026, when the code, database and hosting moved to the client's own accounts with no downtime.

View case study

Agentic OMR for Erkan Ulu Orta Okulu

The challenge: An Istanbul middle school needed answer sheets read from phone photos and scans, across different layouts, lighting and hand-made marks.

What was built: A pipeline in which YOLO finds the answer rows, a custom-trained open-source vision-language model served with vLLM reads the marks, and low-confidence reads go to a teacher for review.

Where it stands: In production, with results flowing into the school portal.

View case study
AI
YOLO
VLM
vLLM
Computer Vision
K-12
Agentic AI
Google Ads
Automation
SEO
Healthcare
Templates
$1M+ Spend
Programmatic

Programmatic Google Ads — Dr. Ayberk Akcay

The challenge: A healthcare practice with many services and locales needed more keywords, ads and landing pages than anyone could keep up by hand.

What was built: An engine on the official Google Ads API that generates campaigns, ad groups, keywords and ads from templates, with bidding rules, budget pacing and change logs.

Where it stands: Active: 2,000+ ads across 1,000+ ad groups, on an account spending over $1M a year.

View case study
Participation Finance
Data Pipeline
Python
PDF Parsing
Next.js
PostgreSQL
JSON API

Sağlama — Source-Linked Participation-Finance Ratios

The challenge: The official participation-finance list is updated once a period while listed companies' balance sheets change every quarter, and investors need to see how each ratio was worked out.

What was built: A free public site, ratio engine and read-only API that recompute three ratios for each measured Borsa İstanbul company from its own financial statements and notes, with the source document, date and method version beside every number. Built solo by Savaş Tutumlu.

Where it stands: Live since July 2026, with 220 companies measured as of October 2026. A forecast frozen before the October 2026 index review matched 238 of 241 calls.

View case study

Ev Çalışması Merkezi — AI Homework Platform

The challenge: A private middle school in Istanbul needed handwritten homework marked and explained question by question, with a whole-class view for teachers.

What was built: An AI homework platform that marks phone photos or PDFs against the teacher's answer key with partial credit, follows mistakes with AI-generated practice questions and lets teachers review the AI's verdicts. A teacher's override always wins.

Where it stands: In use at the school since May 2026; grades 5–8 started the 2026–27 school year on the platform.

View case study
Education
AI
Handwriting
Next.js
FastAPI
Supabase
K-12
Game-Based Learning
WebGPU
iPad
FastAPI
Postgres
Gamification
K-12

Gelgit Adası — A 3D Island Game Where Homework Fuels the Boat

The challenge: Make homework effort, including going back over mistakes, worth something to students, while keeping play time capped and out of school hours.

What was built: A 3D WebGPU island game for students' iPads: right answers, watched solution videos and genuine practice fuel the boat, with a server-authoritative economy, capped play time and no ads, purchases or chat.

Where it stands: In production since 26 Sep 2026 as a teacher-only alpha, two days after the build started. Opening to students is planned.

View case study

Otocropper — Offline Question Cropper

The challenge: Teachers needed scanned question-bank PDFs split into one image per question, named and filed the way the book is printed, on their own Macs.

What was built: A fine-tuned RT-DETRv2 detector paired with Turkish OCR in a notarized SwiftUI app that runs fully offline, with no Terminal, Python or internet connection needed.

Where it stands: 97.1% recall and 96.9% precision at IoU ≥ 0.9 on a held-out 238-page book; in use by the school's teachers since September 2026.

View case study
Computer Vision
RT-DETRv2
PyTorch
OCR
SwiftUI
macOS
K-12