Yevhenii
Budanov

What I built

• TypeScript/Node.js microservices on AWS Lambda + API Gateway with clear service boundaries and independent scaling
• Fully reproducible Terraform infrastructure (RDS, ElastiCache, S3, CloudFront, IAM)
• Custom DB migration runner on ECS Fargate with checksum validation, advisory locking, DynamoDB audit tracking
• Artifact-based deployment with versioned builds in S3, decoupled from cross-repo coupling
• Private VPC architecture with NAT, SSM-based access, and service-to-database isolation
• Redis caching and rate-limiting with explicit key schemas and TTL policies

Architecture

What I built

• Event-driven pipeline (events → feature extraction → inference) with low-latency streaming
• Redis-based feature store for live behavioral signals: timing, sequences, intensity, context-aware metrics
• Gradient-boosted models (XGBoost, LightGBM) for behavioral pattern detection
• Personal-baseline modeling: the system learns each user's signature first, then flags meaningful deviations from THAT
• Synthetic data pipeline for training bootstrap and signal-quality validation
• FastAPI inference service on ECS for scalable real-time predictions
• Multi-level intervention layer (Guardian) with alerts, constraints, adaptive guidance

Architecture

What I built

• Custom Python ingestion from CoinGecko API with retry, rate limiting, idempotent loading
• BigQuery warehouse with partitioned/clustered tables — measured ~11× scan reduction on common analytical queries
• Three-layer ELT (raw → staging → marts) with dimensional modeling: fact + dim tables
• dbt transformations: incremental models, SCD Type 2 snapshots, data-quality tests (not_null, unique, accepted_range, source freshness)
• Airflow DAG: ingestion, transformation, testing, freshness checks with failure alerting
• Terraform for BigQuery datasets, tables, service accounts, IAM
• Dockerized stack: custom Airflow image with dbt, docker-compose for full local environment

Pipeline

What I built

• AI agent architecture for autonomous, multi-step research tasks (LangGraph-based planner)
• Real-time web crawling with intelligent source filtering and provenance tracking
• Hybrid-model inference for faster, bias-minimized analysis
• Structured research reports tailored to analysts, journalists, policymakers, scientists
• Next.js / React frontend with streaming responses
• Python backend with ChromaDB for retrieval; factual-consistency guardrails

Agent loop

What I built

• OCR + geometric layout detection to separate static template from handwritten content
• Region segmentation across multiple test types and variable layouts
• Rule-based and pattern-based grading logic for handwritten answers
• Designed for low-connectivity / offline classroom deployment — no cloud dependency
• Reduced manual bias and ensured consistent evaluation across student submissions

Pipeline

What I built

• Chat-based interface for document query and legal research
• Similarity-based search across massive legal repositories (Llama-2 + ChromaDB embeddings)
• Summarization, citation, and context-aware document retrieval via NLP and LLMs
• Hybrid model routing for compliance-sensitive queries
• Scalable infrastructure with data isolation on dedicated servers
• Iterative refinement experience — attorneys refine queries with follow-up questions

Outcome

• Reduced legal research time per case from hours to minutes
• Improved precision via citation-grounded answers — attorneys see exactly which document and which passage support each claim
• Compliance-safe routing kept regulated queries on approved models
• Data isolation per client meant no cross-firm leakage in shared infrastructure

What I built

• Vector search architecture: indexed thousands of medical documents (PDF, DOC, TXT) with HuggingFace embeddings, ChromaDB for similarity search
• Web-based conversational chatbot interface with contextual follow-ups and dynamic literature filtering
• HIPAA-compliant deployment on dedicated servers with client-specific data isolation
• LangChain orchestration coordinating embeddings, search, and answer generation across models
• Custom medical NLP pipeline: summarization, question rewriting, chunk-level similarity, citation tracking

Outcome

• Healthcare professionals save hours of manual document review per query
• Concise, evidence-based answers with full citations
• Faster diagnosis support, treatment planning, and research workflows
• Compliance-grade isolation eliminated cross-client data leakage risk

What I built

• Designed initial backend architecture (Symfony + PostgreSQL)
• Built REST APIs powering a Next.js frontend with SSR/SSG
• Active backend development and frontend integration through 2022
• Supported JoVE's product family: Journal, Encyclopedia of Experiments, Core, Science Education, Lab Manual, Business, Quiz

Scale

• 1,800+ universities and companies worldwide use the platform
• 30,000+ peer-reviewed videos in the catalog
• Multi-region content delivery with localization in 13 languages
• Used across research, undergraduate education, K-12, and biopharma

What I built

• Re-architected the platform for performance and scale
• Integrated NLP for entity extraction from unstructured data — newsletters, emails, social feeds
• AI-powered search with entity recognition, faceted filtering, and customizable alerts
• REST API for CRM integrations
• Streamlined development workflows and operational practices for long-term efficiency

Context

• Ozmosys is trusted by AmLaw 100 firms and global financial institutions
• Platform aggregates content from LexisNexis, WestLaw, Bloomberg, and other top sources
• Now trained on 1M+ partner-level memos for domain-specific precision
• Modernized intelligence hub that eliminates information overload in high-stakes legal environments