AI Engineering

I build AI features that hold up in production: language models wired into real products, retrieval over company data, agents that use tools safely, and the testing, security and monitoring around them. Alongside this, I have used machine learning in products for forecasting, risk scoring and vehicle valuation since 2018.

1. LLM and model integration

Model choice per task by quality, latency and cost; hosted APIs (OpenAI, Claude, Gemini, Perplexity) or self-hosted Llama on llama.cpp; fallback paths when a provider fails; and cost control per feature. LangChain is useful here as a provider-agnostic layer for models, prompts and output parsing. Seliee lets each customer bring their own provider key or connect a self-hosted model.

2. RAG architecture

Ingestion and chunking of source documents, retrieval and reranking, answers that cite their sources, and refresh jobs that keep the knowledge current. LangChain covers loaders, splitters and retrievers when a pipeline needs them. The personal assistant on this site answers only from my portfolio, and Seliee's in-app assistant loads only the knowledge sections that match the current page and question.

3. Agentic AI

Tool calling and orchestration, state kept outside the model, permissions per tool, and human approval for irreversible actions. LangGraph models an agent as an explicit graph of steps with persisted state, branching, retries and human-in-the-loop checkpoints, and LangChain supplies the tool and model interfaces it builds on. In Seliee's AI Office, agents run scenarios with the panel's own tools and publishing, sending and spending always wait for approval; an MCP server exposes the same tools to external agents.

4. Evaluation and quality

Test sets built from real questions, offline evaluation before release, online checks after it, regression tests for prompt and model changes, and success criteria agreed up front. Seliee's AI Visibility module measures whether OpenAI, Claude, Gemini and Perplexity answers mention a site.

5. Security

Defences against prompt injection and data leakage, tenant isolation, authorization on every tool call, and an audit trail of agent actions.

6. Observability and operations

Tracing, token usage, latency, error rates and cost per feature, monitored in production with the same Graylog and Grafana practice I use on Kubernetes platforms.

Machine learning in products

Forecasting, risk scoring, vehicle valuation and NLP-based translation with TensorFlow, PyTorch and scikit-learn, fed by crawlers and data pipelines.