Igor Tarlinskiy Email & Phone Number

St Petersburg City, Russia

We build LLM-powered assistants that can effectively respond to complex, domain specific technical queries, such as identifying and recommending suitable equipment parameters for purchases from various vendors. Our pipelines is based on RAG, Pre-Rank, Re-Rank, and co.

Moscow, Moscow City, Russia

I have researched LLM acceleration methods that do not require fine-tuning or any external models, offering a plug-and-play solution out of the box. One approach we adopted is LLMA (https://arxiv.org/abs/2304.04487), which leverages the significant overlap between generated tokens via the RAG approach to directly copy-paste some subsequences into the answer.

Moscow, Moscow City, Russia

1. I have developed a semi-supervised approach towards labeling subset of events using GPT-4 model. In the medical background, the patient's medical history and analysis are described using a sequence of ICD-10 codes (International Classification of Diseases, Tenth Revision) and codes specifically representing the patient's conducted analyses - LOINC codes (https://loinc.org). For example, for a patient diagnosed with chronic kidney failure (dis-N18), who then underwent an analysis for the T3/T4 ratio (l-701254) - see https://loinc.org/70152-4, and subsequently diagnosed with type 1 diabetes with kidney issues (dis-E102), the sequence would be "dis-N18 l-701254 dis-E102". That is, for each patient with their history of events, it is necessary to develop a neural embedding to capture the transitional boundaries between clusters of different patients.2. I have developed a model based on the Transformer architecture, in which positional embeddings have been removed, and layers have been modified to reduce the number of parameters to less than 6 million. Trained using triplet loss with a specified margin, the model yielded several unexpected discoveries and, as a result, outperformed the pretrained GloVe model significantly. Visually, the final (2D) clustering allowed for implicit learning of transitive features among clusters, consequently enabling more accurate predictions of transitions from one disease cluster to another.

Moscow City, Russia

AdTech project. We find the best ads for clients and the most relevant audience for advertisers via maximizing ad relevancy metrics so that everyone is happy.

Moscow, Moscow City, Russia

Whenever the client requests help - our mission is to root out the problem that caused the appeal using any NLP techniques we find helpful. The whole pipeline is divided into many sub-problems (e.g. topic modeling, DL classification, semantic search). I'm responsible for all of them to bring the best metrics score.

Moscow, Moscow City, Russia

BERT-based research in generating high-quality sentence embeddings for "Intent recognition" problem.Exploring features across different layers as well as different words resulted in "state-of-the-art" Pearson correlation scoring on "STS 12-16" datasets.

Moscow, Russian Federation

NLP research in an unsupervised approach to build RNN-based models (with the help of auto-encoders) generating high-quality embeddings for both user's query and context key-phrases describing intents.

Moscow, Russian Federation

Backend: (C / C++(-std=c++11))Frontend: (ES5 JS, d3.js).

Moscow, Moscow City, Russia

Explained deep learning ''from scratch'' deriving almost every math equation.From the simplest linear classifiers towards SOTA architectures using PyTorch.The very first lecture is here - https://github.com/xpatronum/NNFromScratch. Let's take a look :=)

Master'S Degree, Mechatronics, Robotics, And Automation Engineering

Model «ATOM» on open-domain neural search and question answering at scale.

Education record

Education record