Aleksandrs Berezņickis

– Architected and led Teaching Voice Assistant leveraging LangChain, LangGraph for agent orchestration and LiveKit for real-time communication, integrated with OpenAI models and Coqui STT/TTS, with a curriculum-aware hybrid RAG architecture. – Led the development of 5+ LLM-based teaching specialist agents built on MCP servers to provide secure, standardized function calling and DB connection interfaces, with an intelligent routing layer that dynamically dispatches queries to the best-fit agent based on context. – Designed and constructed a hybrid RAG architecture for an AI tutor, integrating semantic vector retrieval, curriculum-aware metadata filtering, keyword searching retrieval (BM25) and Sentence-BERT re-ranking to improve answer relevance. – Built a conversational assessment framework for students, incorporating adaptive difficulty and level control in the Chatbot via context engineering and systematic prompt optimization. – Established an enterprise data governance framework for migration upgraded Microsoft Purview with Apache Atlas and Azure AI, improving metadata visibility, data traceability, and security compliance. – Developed the conversational data agent on top of Azure Purview leveraging frameworks like LangGraph and LlamaIndex, enabling users to query table metadata through natural language and Text-to-SQL. – Architected Databricks ETL pipelines to produce actionable data lineage and metadata insights, significantly improving data quality, and implemented batch ingestion to reduce the latency while ensuring the token security policy.

– Led the development and deployment of healthcare voice assistant powered by Multi-component Architecture (ASR, LLM, TTS) using LLM orchestration frameworks like LangChain, LangGraph and Livekit with ReAct and Chain-of-Thought method, utilizing the Hybrid RAG. – Fine-tuned over 10 open LLMs (Llama, Qwen) for adopting own healthcare datasets and conversational AI using QLoRA on Amazon SageMaker, reducing memory usage up to 90%, speed up 3x training duration. – Utilizing pre-retrieval techniques like chunking strategies, metadata extraction and query optimization to construct Hybrid RAG leveraging PostgreSQL and FAISS, boosted retriever performance through integrating ColBERT re-ranker as a post-retrieval to reduce non-grounded (hallucinated) responses. – Mitigated hallucination and toxicity risks through RLHF informed by 200+ expert annotators, strengthened factuality via mandatory source citation, and deployed Guardrails AI and TinyBERT as guardrails to enforce HIPAA/GDPR and PHI compliance. – Optimized the LLM serving pipeline by integrating vLLM, PagedAttention and TensorRT-LLM with dynamic batching on GCP, resulting in a 30%+ higher GPU utilization and a 0.8s reduction in time-to-first-token (TTFT). – Defined and led the strategy for LLM optimization, leveraging smart model routing, Redis-based caching, prompt optimization and deep system-level diagnostics to eliminate performance bottlenecks across production AI platform and reduce the token usage and increase the throughput. – Architected component-based evaluation pipeline combining traditional metrics (BLEU, BERTScore) and LLM-as-a-Judge, enabling robust quality monitoring via MLflow and LangSmith, and accelerating prompt iteration, strengthening agentic workflow reliability.

– Fine-tuned 5+ LLMs using LoRA, model selection for efficiently extracting and summarizing information from extensive document sets, while effectively removing jargon and irrelevant contents, optimized OCR through model optimization (like pruning, quantization), reducing latency by 35% significantly enhancing real-time processing. – Developed a Name Entity Recognition (NER) system for e-commerce, achieving an F1 score of 95% for entity identification, improving 2% revenue growth. – Created a sentiment analysis engine that processes 8,000 daily customer reviews using Tensorflow, Spacy and NLTK, achieving 95% accuracy in product development and achieving ROUGE-1 score 41 for 35,000 product description per day. – Implemented a masked face recognition system using transfer learning while freezing the original model, keeping up optimal performance 97%. – Designed and implemented an end-to-end MLOps pipeline using Kubeflow, integrating CI/CD process using Git and Docker containerization reducing model deployment time 30%.

- Worked on production-grade XGBoost models for AlphaLens; co-led feature engineering and validation pipelines, contributing to ~18-25% accuracy gain over linear/logistic baselines. - Optimized model serving latency by ~30% through quantization and dependency reduction in lightweight inference containers, working alongside core model development. - Designed cross-validation frameworks and automated feature importance tracking to enable reproducible model improvements and faster iteration cycles. - Migrated scheduled batch pipelines to near-real-time micro-batches, improving data freshness from daily to sub-hourly and increasing throughput by 3–5x. - Reduced pipeline failures by ~60% by implementing idempotent writes, dead-letter queues, and automated retry logic with exponential backoff in Airflow DAGs. - Built preprocessing and postprocessing modules (audio normalization, punctuation restoration, diarization stitching) that reduced manual transcription effort by ~70%.