Daniel Garayan Email & Phone Number

I conduct research in the field of artificial intelligence, write technical specifications and define project tasks, as well as develop and implement prototypes/production.

Москва, Москва, Россия

Build LLM based product. Did all phases from system design to inferencing

Мсида, Мальта

• Enhanced internal chatbots by enabling summarization and comparisons of related PDF documents, utilizing retrieval-augmented generation (RAG) through LangChain, LangSmith and Qdrant vector databases.• Fine-tuned Mistral 7b and Llama2 13b models using PEFT Oobabooga tools and methods, incorporating LORA and RLHF techniques for enhanced performance.• Designed and implemented system for seamless text-to-SQL query processing, optimizing data retrieval and enhancing overall system efficiency.• Vector search based news recommendation system• NLP based Robotic process automation (RPA)• Participated in code reviews, providing feedback to team members and ensuring code quality and best practices.• Educated internal stakeholders through presentations to over 30 colleagues on the latest in Generative AI𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡:• Conducted comprehensive benchmarking assessments on leading vector databases, including PgVector, Qdrant, Chrona, Milvus, and Pinecone, to optimize performance and drive data-driven decision-making.• Conducted model comparisons with Hugging Face for efficient embedding selection. ( bge base, text-embedding-ada-002 and etc.

• 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭 𝐃𝐞𝐭𝐞𝐜𝐭𝐢𝐨𝐧 𝐢𝐧 𝐕𝐢𝐝𝐞𝐨𝐬Highlight detection models are trained to recognize features that make visual content appealing or interesting to a wide audience. The primary objective is to identify these moments in videos and subsequently create vertical videos.Stack: Audio tagging, Sentiment analysis, FFmpeg, openAI api, FastApi .and etc.• Backend tasks• Set up code review and delivery process

𝐈𝐦𝐚𝐠𝐞 𝐍𝐒𝐅𝐖 𝐂𝐥𝐚𝐬𝐬𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧• Developed a pipeline for collecting and processing a custom image dataset. Implemented a hash-based mechanism to eliminate similar photos• Used ResNet for training the NSFW classification model.• Integrated the model into a Node.js backend.• Ensured data quality by manually reviewing and marking images.• Conducted testing and ongoing optimization for model accuracy.• Collaborated with teams, documented processes.

The software is designed for quick mystery box purchases from binans. The program was built into Telegram for the convenience of buyers.

Machine Learning Engineer

Worked with US and EU clients to develop Python solutions for web development, data processing, automation, and system integration.

Москва, Москва, Россия

Real-time processing of the data of the object touching the surface from the camera. Transferring the data to the server via the TUIO protocol. After which the data was applied in Unity3D.