Farid H.

• Fine-tuned CosyVoice TTS models on proprietary datasets for 30 distinct speakers, improving voice quality and speaker consistency
• Built a scalable inference service using Ray and FastAPI for low-latency text-to-speech generation
• Implemented Speaker diarization pipelines for long-form audio, enabling accurate segmentation and extraction of individual speaker voices
• Conducted experimental work on speech-to-speech systems by fine-tuning the Moshi pipeline; replaced base model (Helium) with Qwen3 for improved flexibility (project in progress)
• Fine-tuned Qwen Image Edit on custom paired (before/after) datasets for image editing
• Fine-tuned text-to-image diffusion models (FLUX, Stable Diffusion XL) and implemented DreamBooth for subject-driven generation using Hugging Face

TTEK Inc. focuses on creating and implementing innovative technologies to enhance the efficiency of customs and border processing.

• Used Large Language models with advanced prompting techniques for text to sql. Created one LLM agent in order to do entity extraction. Then used output of that agent in SQL agent in order to create an sql query.
• Utilized advanced Prompting techniques with LLMs (gpt4-o and Claude) to extract valuable insights from diverse data sources in order to create risk rules in customs.
• Created an LLM Agent using an LLM (gpt4-o) together with serper api in order to predict the NAICS code for a given company name.
• Used LLMs for price prediction.
• Designed a dual-head classification model leveraging the BERT LLM architecture for precise categorization of products into Harmonized System (HS) codes.
• Built a Named Entity Recognition (NER) model to extract product names from descriptions, using BERT based LLM for token classification.

• Engageret i forsknings- og udviklingsaktiviteter.
• Udviklede en REST-server til Autocorrect-modulet, der giver problemfri og effektiv funktionalitet.
• Benyttede OPTUNA til automatisk justering af hyperparameter, der optimerer ydeevnen af modeller og algoritmer.
• Anvendte forklarlige AI-teknikker, herunder værktøjer som Captum, Lime og SHAP, for at give fortolkelig indsigt i modelforudsigelser og beslutningsprocesser.
• Udførte finjustering af BERT-modeller for hensigtsklassificeringsopgaver og forfinede deres præstationer på specifikke kontekstuelle forståelsesopgaver.
• Implementerede en PyTorch-baseret DIET-klassificering fra bunden, der muliggør bred hensigtsgenkendelse og klassificering.
• Udførte eksperimenter med RASA's open source konversationelle AI-ramme og udforskede dets muligheder og potentielle programmer.
• Udforskede end-to-end-dialogmodeller ved hjælp af sprogmodeller (LLM'er) og undersøgte deres effektivitet og ydeevne i AI-samtaleopgaver.

Voiceloft, an Azerbaijan-based platform that promotes the development of speech recognition technologies by collecting voice data.

• Fully responsible for the ML projects;

• Working on ASR model and leading 2 interns on NLP tasks;

• Fine-tuned Wav2Vec2 baseline model with our dataset.

• Har arbejdet med forskning og udvikling samt dataanalyse;
• Anvendte sentiment analysemodeller fra start til implementering, ved hjælp af Python, PyTorch, scikit-learn, Docker og Flask.

• Havde det fulde ansvar for at føre tilsyn med og styre hele livscyklussen for ML-projekter.
• Succesfuld implementering af en avanceret aserbajdsjansk tale-til-tekst-model ved hjælp af Kaldi-toolkittet og Python.
• Udviklede en aserbajdsjansk tekst-til-tale-model ved hjælp af Tacotron2 og parallel vegansk model til syntese i PyTorch.

• Fully responsible for the ML projects and crawlers;

• Applied Generic Crawler to get data from any news page, even if they change their design;

• Built the ETL process, getting news data from 200 different Azerbaijani news channels;

• Applied Similarity Detection, Text Categorization, NER, Sentiment;

• Analysis, Keyword Extraction, Summarization to those data;

• Finished Projects: aqreqator.az.