Mir Siadaty

• Architecting and deploying end-to-end machine-learning solutions on vendor clouds (AWS, Azure, etc). • Creating generative neural nets for textual, visual, audio, and structured data, serving real-time predictive services. • Text-processing on natural languages (NLP), biomedical texts (UMLS), and machine languages (automation). • Building data science teams, deploying full-cycle ML-platforms, implementing real-world business solutions. ◦ RAG, via Meta’s Llama-v2-70Billion LLM, also via OpenAI’s ChatGPT and GPT4 ◦ Fine-tuning LLMs via TRL - Transformer Reinforcement Learning, and PEFT - Parameter-Efficient Fine-Tuning, as well as merging LoRA adapters back into the original LLM. ◦ LLM agent, via GPT4, with access to tools like Google API, python code interpreter, and ‘self-ask+search’ prompting ◦ RAQA via HuggingFace + FAISS + LangChain; LLMs "mistralai/Mistral-7B-Instruct-v0.1" and "sentence-transformers/all-MiniLM-l6-v2" ◦ Automated evaluation of RAG pipeline, via up-to-8 quantitative metrics (context precision, context recall, context relevancy, context entity recall, faithfulness/groundedness, answer relevancy, answer semantic similarity, answer correctness), using Ragas + Langchain + OAI GPT4 / Meta Llama3-70B-instruct ◦ LLM-based unsupervised text-clustering, using OAI GPT4 as well as Meta’s Llama3-70B-instruct ◦ Topic modeling, via BERTopic (SentenceTransformer, Umap, Hdbscan, KeyBERT, openhermes-2.5-mistral-7b), as well as pure LLM topic-modeling via goliath-120b.Q2_K.gguf (serving by llama_cpp) ◦ Sentiment analysis, via Zephyr-7Billion LLM, on Ollama ◦ NLP, such as automated knowledge-graph extraction, via spaCy/textacy as well as Mixtral 45Billion LLM, also Babelscape/rebel-large 406M ◦ Image generation, via Stability AI’s Stable-Diffusion XL 3.5Billion U-Net Transformer model ◦ Human natural language creation/translation, 200 languages, via Meta’s NLLB-MOE 54Billion LLM, also SeamlessM4T 2.3Billion NN text and speech generation ◦ Sparse-vector-space text classification, via Sklearn’s n-gram generation, Tfidf Vectorizer, and Xgboost.