Dmytro B.

Project overview:

Developed an image-based product search system for a major retail chain with 100K+ SKU catalog. Customers upload a product photo and instantly find matching items in stores. The solution uses a novel approach: LLM generates semantic descriptions from images, which are embedded and matched against the catalog via vector similarity search.

Responsibilities:

• Designed AI pipeline: image → Gemini Vision description → text embedding → Vertex AI Matching Engine vector search (top-50, cosine ≥ 0.60)
• Implemented deterministic prompt engineering for consistent product descriptions optimized for embedding quality
• Built smart sync strategy with automatic incremental vs full rebuild selection based on change volume
• Developed role-based JWT auth (Owner/Admin/User) with domain restrictions
• Deployed with Docker Compose, Nginx, automated SSL

Achievements:

Sub-second visual search across 100K+ products. Smart adaptive indexing automatically selects between incremental and full catalog rebuilds, keeping the index up to date with minimal downtime.

Technology stack:

Project overview:

Built an AI-powered knowledge base for a large logistics company group (5 subsidiaries). The system syncs thousands of PDFs from Google Drive, indexes them in Vertex AI Search, and provides employees with conversational AI search - delivering answers with citations from corporate documentation.

Responsibilities:

• Architected multi-tenant RAG system with per-company datastores, buckets, and Drive folder mappings
• Implemented full document lifecycle: Drive → GCS → Discovery Engine with incremental sync and deletion cleanup
• Built Celery Beat periodic sync with retry logic and repository pattern
• Developed conversational search with persistent sessions and structured citation extraction
• Implemented multi-language detection (9 languages) to override Discovery Engine's auto-detection

Achievements:

Manages 6 isolated datastores across 5 companies with 10 Drive source folders. Multi-turn conversations with context retention. Parallel async API calls reduce first-query latency significantly.

Technology stack:

Project overview:

Built a distributed ETL platform for processing massive email archives (EML files in ZIP/7Z/RAR) into structured PDFs, uploading to Cloud Storage, and indexing in Vertex AI Search. 9 isolated workspaces with strict data separation - designed for a government organization.

Responsibilities:

• Designed distributed task architecture: 9 isolated queues across 3 Celery workers with chord-based batch processing
• Built EML pipeline: archive extraction → HTML parsing → format conversion via LibreOffice → PDF generation
• Implemented real-time status: Redis pub/sub → FastAPI → WebSocket broadcast
• Configured OAuth2 proxy with email allowlist for access control
• Built retry-aware GCS upload and Discovery Engine incremental import

Achievements:

18 parallel workers across 3 containers with per-app queue isolation. Real-time WebSocket status updates. Batched operations with retry logic processing thousands of email files.

Technology stack: