Viktor Veselov

Built an LLM evaluation framework using Langfuse tracing to detect failure modes and enable continuous integration testing. Deployed Playwright E2E framework, establishing CI/CD coverage and reducing cron job processing time by 85%.

Architected KB as IaC, automating RAG deployment per-user via S3 vector store, eliminating manual environment configuration. Co-designed a tool registry with dynamic prompt generation and a centralized control panel for feature-flagging and agent permissioning, reducing token costs by 30% and enabling paid subscription tier management across dev and prod.

Trained a custom BERT model on 100+ synthetic examples to resolve document versioning conflicts across enterprise folders, normalizing language and generating embeddings for semantic similarity in RAG. Reduced agentic workflow from 5 agents to 3, cutting operational costs while maintaining output quality.

Evaluated and ranked outputs across multiple LLMs, engineering prompts to expose failure modes and improve model reasoning quality as part of Anthropic/OpenAI/Meta RLHF pipelines.
Tech Stack: Python, JavaScript, Docker, SQL, XGBoost.

Prototyped model reliability developing advanced computer vision systems for light distribution.
Prototyped an RGB image-based body temperature system combined with a custom Computer Vision model, with approximate 60% accuracy in measuring body temperature.
Tech Stack: Python, Jira, Computer Vision, CNN, Resnet, XGBoost, YOLOv8, Roberto, BERT, Vector Embeddings, RAG

Prototyped a pipeline to reduce data storage costs by approximately 10%. Developed an evaluation pipeline and automated the migration of 1.2 petabytes of data using advanced LLM solutions on AWS.
Tech Stack: Python, AWS, BERT, AWS Lambda, API Gateway.