Bohdan – Python, LangChain, LLM

An AI-powered platform for automated extraction and reconciliation of financial data from private equity documents. The product combines vision-language models, document matching, and an interactive PDF review interface to help users extract, compare, and validate financial data efficiently.

• Fine-tuned Qwen2.5-VL on financial documents for field-level extraction from FS and CAS PDFs and integrated a specialised LP extraction model with a RAG embedder for page search;
• Built an async GPU inference pipeline with multi-model architecture, CUDA memory optimisation, device-agnostic fallbacks, and request-based cancellation support;
• Designed a series-aware fuzzy matching algorithm for intelligent FS/CAS document pairing to prevent false matches between fund versions;
• Implemented an automated reconciliation engine with deviation calculation, NAV/commitment basis selection, deviation flagging, and Excel backtesting export;
• Built a React/TypeScript frontend with side-by-side PDF viewing, bounding box highlighting, drag-to-select field correction, and batch comparison;
• Architected a FastAPI backend with SQLite persistence, batch ZIP processing, real-time status polling, and full-state undo/redo support.

An AI-powered tax research agent with hybrid search capabilities designed to improve the quality and relevance of retrieved information from large document collections.

• Built a Pinecone hybrid search module and integrated it into the backend architecture to improve retrieval metrics;
• Created an additional sparse vector index for hybrid search;
• Reindexed large document databases with sparse vectors;
• Performed deployment and testing.

An AI-powered veterinary assistant that combines computer vision, NLP, and multi-agent workflows to support pet healthcare. The product provides conversational guidance, veterinary Q&A, health information gathering, and image-based disease analysis through a unified chatbot experience.

• Designed a multi-agent LangGraph architecture with persistent conversation memory, observability, and S3-backed consultation history;
• Built specialised agents for veterinary diagnosis, pet health gathering, and Q&A with multilingual support;
• Deployed a multi-agent expert network with specialised veterinary personas and MongoDB-backed pet medical record access;
• Integrated a vision pipeline for pet disease detection, segmentation, and batch image analysis into the chatbot diagnostic flow;
• Architected a scalable FastAPI backend with Docker, Nginx, Celery, Firebase, WebSocket streaming, JWT auth, and CI/CD on AWS;
• Researched MCP and evaluated multiple frontier models for veterinary domain use cases.

An ML-based bid price optimisation system for a high-volume DSP using real-time CTR/CVR prediction under strict auction latency constraints. The system optimises bid pricing and execution by combining predictive models, feature serving, and automated model deployment.

• Designed a CTR and CVR to eCPM bid calculation pipeline with bid shading logic for first-price auction optimisation;
• Quantized XGBoost models to ONNX and deployed them on SageMaker multi-model endpoints to reduce p99 inference latency below 10ms;
• Pre-aggregated user and publisher features into SageMaker Feature Store via Kinesis upstream;
• Built an ElastiCache hot layer for sub-millisecond user profile reads at peak throughput;
• Automated daily retraining via SageMaker Pipelines with shadow evaluation before endpoint promotion and used AWS Lambda for bid routing and budget pacing.