Enes – Machine learning, Python, NLP

Assistant AI chatbots for students in an online school for all grade levels.

• Created assistant AI chatbots for students in K12;
• led the AI team;
• used vector databases and RAG;
• used open-source LLMs and GPT API;
• worked on the API integration;
• was responsible for all the back-end work.

The tool that helps authors generate fictional prose.

• Used SFT, DPO and GRPO for better fictional prose with the authors' own writing style;
• used Reinforcement Learning.

LinkedIn-like candidate recommendation engine for C-Level people (board members, CTO, CFO, etc.)

• Was executive for a search recommendation engine with LLMs;
• created job function and specialism prediction from candidate info, etc.;
• used open-source LLMs and GPT APIs;
• finetuned Mistral-7B and Mistral-8x7b in single and multi-GPU settings;
• used quantization methods such as GPTQ, AWQ.

Generative AI system to create advertisement text and images using LLMs and text2image models.

• Worked on OSS and API based LLM such as llama2 and OpenAI GPT APIs;
• fine-tuned and used several text2image models (SD versions, SD, XL, etc.);
• created APIs for all these using flask API;
• used chatbots in advertising to create advertisement text and paired them with text2image models to create banners.

Categorization and processing of Insurance documents with AI.

• Created several models for insurance doc type classification for different file formats and handwritten scanned docs;
• fine-tuned and used several OSS deep learning models such as pix2struct, donut and layoutllm to process the files;
• created APIs on top of these models;
• used chatbots for insurance documents understanding.

Vehicle make-model-body type-color-year prediction ML model to make users' car ads creation process much easier.

• Worked on a Very Large-Scale NLP-Computer Vision Project to infer categories of ads (trained and tested with tens of millions of ads) and deployed successfully;
• worked on a Very Large-Scale Machine Learning (Deep Learning-Computer Vision) Project (Fine-grained vehicle classification: make, model, body type, color, year of a car). Designed, implemented, and deployed it end to end, including:

1. Prototyped in Keras, implemented with Tensorflow with high training performance;
2. Trained with more than 32+ millions of images on Google Cloud Compute Engine in multi-GPU setting, reaching state-of-the-art accuracy Deployed the model on Google ML Engine and exposed it on Compute Engine with Python Flask;
3. Deployed the model on Tensorflow Serving on on-premise datacenter and implemented a Rest API with Flask for production use.