AI & ML interests
E-Commerce
Recent Activity
AIPROJE
Turkish datasets, domain-adapted AI models, and a decade of applied engineering ā e-commerce is our flagship business, not our whole story.
Who we are
AIPROJE was founded as an AI and software projects company, not just an e-commerce infrastructure provider. With a decade of handsāon experience across AI/ML, software engineering and systems integration, we started by building AIāpowered eācommerce solutions and have since expanded into developing Turkish language models, datasets and domaināspecific AI for multiple industries. Our platform, aiproje.com, is a comprehensive eācommerce infrastructure for Turkish B2B, B2C and dropshipping sellers ā supporting multiālanguage and multiācurrency operations ā but our work does not stop there. We also create AIāpowered mobile applications, visual editing tools and tailored AI solutions for sectors ranging from automotive to logistics.
This Hugging Face organization is where we publish our AI research and engineering output:
- Build Turkish datasets, including Turkish reasoning datasets with step-by-step (
think) chains rather than plain answers - Fineātune open models for concrete, domaināspecific jobs instead of generic chat
- Optimize inference and adapt models to run on modest hardware ā quantization, LoRA, lowāVRAM deployment
- Work through the decisions behind base model development ā architecture, training, and adaptation choices ā and share what we learn along the way
- Feed all of the above back into our commercial products: AIāassisted product search, translation, image enhancement, pricing, and entirely new AIādriven applications
Leadership
AIPROJE is led by DoÄukan Atakul, who also founded the Bursa AI Community (Bursa Yapay Zeka TopluluÄu). His background covers sixāplus years across AI/ML systems, computer vision, model compression and fineātuning, lowāresource and edge deployment, domaināspecific model development, and systems integration ā alongside the broader software engineering, Linux, and server/network work that the AIPROJE platform itself is built on.
He also publishes Turkishālanguage technical videos on his YouTube channel, covering LLMs, fineātuning, quantization, MoE, edge AI, and tools like CUDA, PyTorch, TensorRT, vLLM, and Ollama ā aimed at explaining AI as engineering rather than hype, and growing the amount of serious Turkishālanguage technical content out there.
What's here
Our first release wave centers on Turkish automotive eācommerce ā spareāparts search, compatibility checks, and technical support ā a concrete domain we're using to prove out our approach before generalizing to others.
š¤ Models
| Model | Base | Task | What it does |
|---|---|---|---|
qwen3-1.7b-otomobil-yedekparca-lora |
Qwen3-1.7B (Unsloth, 4ābit) | Text generation Ā· LoRA | A Turkish automotive spareāparts assistant covering product lookup, pricing, compatibility checks, troubleshooting, function calling, and multiāturn, stepābyāstep reasoning. Fineātuned on tens of thousands of synthetic reasoning QA pairs. By our own benchmark notes it's solid at product search and function calling, weaker at compatibility checks and openāended chat ā an experimental checkpoint testing how far a small 1.7B model can go in production eācommerce. |
kumru-2b-flirt-lora |
vngrs-ai/Kumru-2B | Text generation Ā· LoRA | A small sideāexperiment in shifting a Turkish base model's conversational tone toward casual, informal chat. Purely a research/fun project ā not intended for production or professional use. |
Both models are released under Apache 2.0.
š Datasets
| Dataset | Size | License | What's in it |
|---|---|---|---|
turkce-otomobil-yedek-parca-soru-cevap-akil-yurutme |
~11,454 rows | MIT | Our largest release: Turkish reasoning data for real spareāparts compatibility questions ā OEM codes, brands, and model years across dozens of manufacturers ā each with a think column showing the reasoning behind the answer. |
turkce-sentetik-otomotiv-yedek-parca-akil-yurutme |
~990 rows | MIT | Synthetic Turkish reasoning data on part identification, compatibility, and supportāstyle conversation, with a deliberate mix of tones and customer personas. |
turkce-otomotiv-yedek-parca-youtube-soru-cevap-akil-yurutme |
~260 rows | MIT | Reasoning Q&A distilled from Turkish automotive YouTube content, ranging from maintenance and diagnostics to adjacent engineering topics. |
turkce-otomobil-bakim-soru-cevap |
1.69 MB | MIT | A more general Turkish Q&A set on car maintenance and common faults, without the reasoningāchain annotations of the other three. |
All four are Turkishālanguage and MITālicensed. Combined, the three reasoning sets alone total roughly 12,700 rows ā and more reasoning data and checkpoints are on the way.
What we build
Our flagship product, AIPROJE, is a modular eācommerce platform for retail, wholesale and dropshipping sellers, including exporters. It handles order management, invoicing, stock, and reporting, with multiālanguage/multiācurrency support and integrations for local carriers, marketplaces, and payment providers ā plus AI features (automatic translation, image enhancement, product copy, marketābased pricing) built directly into the product.
Beyond eācommerce, we develop AIāpowered mobile applications, image and video editing tools, and domaināspecific AI solutions for industries like automotive, logistics, and retail. These projects often involve creating custom datasets, fineātuning models, and deploying optimized inference on edge devices ā contributing directly to the Turkish AI ecosystem.
Get in touch
We're glad to hear from researchers, developers, or businesses working on Turkish NLP, eācommerce AI, or anything adjacent.
- š Website ā aiproje.com
- š» GitHub ā @aiproje
- š„ YouTube ā @dogukanatakul
- š¦ X ā @TlCARET
- š¼ LinkedIn ā company/aiproje
- šø Instagram ā @aiproje
- āļø Email ā info@aiproje.com