Bogdan – AWS, Python, LLM

A multi-agent platform that automates legacy code migration, such as moving SQL stored procedures to C#, on codebases of over 100K lines.

• Led a team of 4 building a multi-agent system that migrated legacy code, for example SQL stored procedures to C#, on projects over 100K lines.
• Built a 12-agent system with separate roles under one coordinating agent.
• Had agents pull code context through MCP servers and compare screenshots of the old and new UI through browser automation.
• Ran migrations in parallel on AWS as containers.

A customs platform for UK import declarations. It extracts data from invoices, packing lists and certificates and submits declarations to HMRC for customs brokers.

• Worked directly with UK customs brokers to learn the import declaration process and turned their daily workflow into product decisions and customs rules kept as data, not code.
• Contributed as the main contributor to the first production version.
• Rebuilt the platform as a TypeScript monorepo.
• Built the document extraction pipeline for invoices, packing lists, and certificates using local PDF parsing, vision OCR, and LLM extraction.
• Integrated HMRC CDS submissions and notifications, NCTS transit, IPAFFS, and the UK Trade Tariff API.
• Guarded releases with replay-based validation, automated tests, and browser scenarios filing to the HMRC sandbox.
• Deployed the platform on Azure using AKS and Helm.

A product feature that automatically splits IT tickets into solvable subtasks using LLM reasoning approaches.

• Automated splitting IT tickets into solvable subtasks.
• Benchmarked CoT, Tree of Thoughts, LATS, ReAct, Reflexion, and multi-agent setups on shared metrics.
• Shipped the best-performing approaches into the product.

A legal search engine built on RAG over 400,000 documents, returning relevant passages and a plain-language answer in 10–15 seconds.

• Built a RAG search engine over 400,000 legal documents that returned relevant passages and a plain-language answer in 10-15 seconds.

An LLM assistant that routes each query to a specialised agent able to run calculations, call pricing APIs, or write and run SQL.

• Built the request router for the xPortal assistant.
• Routed each query type to its own LLM agent that could run calculations, call pricing APIs, or write and run SQL.

A WhatsApp booking agent for a platform serving 700 partner salons and 180,000 active users, taking over appointments previously booked by phone.

• Led a small team that took a WhatsApp booking agent from concept to production for a platform serving 700 partner salons and 180,000 active users.
• Built an agent that booked over 10 appointments a day that used to be handled by phone.

Machine learning work including a named entity recognition solution taken to production and model optimisation for faster inference.

• Cut inference time and model size by up to 10x with quantization, pruning, and compilation, without losing accuracy.
• Took a named entity recognition project from concept to production, reaching 98% F1 on the main classes.

Research and development of deep learning methods for cardiac MRI, including segmentation, classification and synthetic data generation.

• Worked on 7+ research projects applying deep learning to cardiac MRI.
• Produced patent filings and a first-author journal paper through the research work.
• Built 3 annotation tools for time-series and 4D medical data used to label thousands of samples.
• Applied 2D/3D segmentation, regression on 3D+t MRI, multi-modal classification, GAN-based synthetic data, few-shot learning, and reinforcement learning for landmark detection.

An interactive tabletop augmented reality product that combines machine learning and blockchain-based image sourcing.

• Built keypoint regression and object detection on RGB-D streams for a smart lamp.
• Built an automated pipeline for collecting and labelling data.

Computer vision research on replacing SIFT feature matching in medical images with deep learning.

• Redesigned SIFT feature matching for medical images with deep learning.
• Presented a poster at the EEML summer school.