My AI work is native to the platform, not bolted onto it. Models train, deploy, and run inside the same .NET services as everything else — the same authentication, logging, and release pipeline the rest of the system already passes through. Classical ML where it is explainable and cheap, language models where they genuinely earn their place, and an audit trail either way. That is what makes AI shippable in banking and government.
ML.NET in production
Native machine learning for .NET applications — models trained, versioned, and deployed inside the same ASP.NET Core services they serve, with no separate Python runtime to operate or secure.
Matching & ranking engines
The candidate-to-job matcher behind the Intelligent Recruitment System, scoring on skills, experience, and requirements — it cut screening time by 40% and removed 18+ hours of manual review a month.
Classical ML, correctly applied
Regression, classification, clustering, and anomaly detection — chosen because they are explainable, cheap to run, and defensible to an auditor, which matters more than novelty in regulated systems.
Predictive models & data analysis
Turning the transaction and operational data these platforms already hold into forecasts and signals, without moving it outside the estate that governs it.
Intelligent automation
Workflow automation with models in the loop — screening, classification, and routing that eliminate repetitive back-office work rather than merely tracking it.
AI inside business platforms
Integrating models into existing .NET estates so inference sits behind the same authentication, logging, and release pipeline as every other endpoint — not in a side system nobody owns.
LLM integration in .NET
Azure OpenAI and Semantic Kernel wired into ASP.NET Core services, with tool calling and structured outputs treated as typed contracts rather than string parsing.
Model Context Protocol
Connecting models to real systems through MCP — typed tool definitions and controlled context, so an assistant can act on platform data through one governed interface instead of bespoke glue per integration.
Retrieval over systems of record
Embeddings and vector search layered onto the MSSQL, PostgreSQL, and MongoDB estates that already exist, so answers cite the organisation’s own data instead of the model’s memory.
Event-driven inference
Model calls run as consumers on RabbitMQ or Kafka rather than blocking an HTTP request — with retries, dead-letter queues, and backpressure, the same way every other integration is built.
Guardrails & auditability
Full prompt and response audit trails, role-based access on AI endpoints, and a human in the loop on anything that moves money — so an AI decision can be reconstructed months later.
Evaluation & cost control
Golden datasets and regression tests so a model or prompt change cannot silently degrade output, plus caching, token budgeting, and model routing to keep spend predictable.
