Experiences
- Promoted the adoption of Model Context Protocol (MCP), building FastMCP servers over internal resources and a multi-agent demo platform (A2A agents, a discovery registry, an LLM orchestrator planning over discovered capabilities)
- Re-designed and built an LLM extraction pipeline replacing multiple chained LLM model calls with deterministic pre-filtering, embedding-based page localisation with FAISS, per-template structured output and the evaluation harness against ground truth
- Built the provider-abstraction layer and self-hosted model support (Ollama) for an internal RAG platform, enabling operation entirely on-machine for data that cannot leave the perimeter
- Delivered event-driven services on Azure Durable Functions, a document-processing pipeline using Service Bus and external-event orchestration, and an AML screening service with fan-out/fan-in and full audit history
- Introduced Software Engineering practices on a legacy product: pre-commit hooks, CI checks, code review, issue tracking and mentored junior developers
- Took the Deep Learning models from research artefact to production asset: MLflow experiment tracking and model registry, with the serving application loading models directly from the registry, and monitoring of live prediction confidence and distribution
- Turned a data-science proof-of-concept into a production document-extraction service (FastAPI, OCR, models versioned in S3); diagnosed unbounded memory growth and re-architected it into a queue-and-scheduled-drain design with a persisted job state machine under GDPR retention constraints
- Designed and built an AML/KYC risk-scoring platform still in production today: a versioned Python scoring library with YAML-driven regulatory weights, and the FastAPI service on top, deployed on Kubernetes across three environments
- Built the cleaning and normalisation layer of an AML platform on Databricks (dbt, medallion architecture) over billions of transactions, and an NLP platform extracting business concepts from operational-loss descriptions
- Core contributor to a vehicle-damage assessment system that became a commercial product, still sold and maintained today: from photographs, a pipeline of four Deep Learning models detects the vehicle, classifies and segments damage and car components, intersecting the masks to localise damage per part
- Built the training dataset from scratch (scraping, annotation schema, CVAT labelling over real insurance claims) and corrected a train/serve distribution mismatch through re-labelling and targeted data augmentation; trained and benchmarked classification, object detection and segmentation deep learning models (DenseNet, YOLOv3, Mask R-CNN)
- Built distributed ETL pipelines and analytical tooling in Scala, Python and Spark for large-scale insurance processes, and containerised an Elasticsearch / Logstash / Kibana stack on the internal on-premise environment