AI Expert

What does an AI expert do?

• Chooses the right problem. Not every job needs AI; sometimes a simple rule engine is cheaper and more reliable. The expert distinguishes where a model genuinely creates value.
• Prepares the data. Collects, cleans and chunks the sources the model will answer from, and sets up access permissions.
• Sets up a RAG architecture. Instead of making the model memorise company data, it lets the model answer by fetching the right document at the right time.
• Writes prompts and evaluation. Builds test sets that measure answer correctness; does not settle for "looks good".
• Manages cost and latency. Lowers usage cost with model routing, caching, smaller models and streaming responses.
• Sets security and boundaries. Access control, personal data masking, human approval and abuse prevention.
• Monitors in production. Continuously tracks answer quality, error rate, latency and cost metrics.

What are LLM, RAG and agents?

LLM (large language model) is a model trained on a large amount of text that understands and generates text. On its own it does not know your company's current data; that is why it is not enough by itself in most enterprise use.

RAG (Retrieval-Augmented Generation) is the model fetching relevant documents from a retrieval layer before generating an answer. This way the answer is based on your data and can cite sources. This is what matters in areas such as policies, product documentation, contracts and patient information.

An agent is a system that makes decisions step by step and can use tools (search, database, files, API) to reach a goal. Agents are powerful, but they should not go to production without an authority boundary, an approval point and an audit trail.

When does AI make sense, and when not?

• Makes sense: free-text understanding, summarisation, classification, finding answers inside documents, customer support, content generation and multi-step workflows.
• Be careful: exact calculation, decisions with legal consequences, high-risk medical guidance and unmonitored automated actions.
• May not be needed: work solvable with rules, areas with no data, and small one-off tasks.

The real problems of AI in production

• Hallucination. The model can answer wrongly while sounding confident; RAG and source citation reduce this but do not eliminate it.
• Cost. Token cost grows as usage grows; model routing and caching are essential.
• Latency. User experience requires streaming responses, precomputation and the right model choice.
• Privacy. Personal or sensitive data must be masked and the provider chosen before it is sent to a model.
• Evaluation. "Works well on examples" is not enough; a measurable test set and regular checks are needed.
• Model changes. Providers update models; the system should not be locked to a single provider.

My experience

KaiMoi is the AI agent and backend product I built. I designed a multi-agent runtime made up of the Kai (autonomous developer) and Moi (reviewer and orchestrator) roles. The system does cost-sensitive routing between DeepSeek, Claude and Anthropic. I am also developing an open-source multi-agent orchestration framework called Cogentic.

At DoktorAsistan I built LLM- and RAG-based question-and-answer, content generation and assistant flows for patient and doctor communication. Multi-channel notifications, automated email flows, CRM enrichment and analytics run in the same system.

I also assessed whether AI can fully manage a site or product in my article websites managed with AI.

How does the process work?

1. Discovery. Which job, with which data, and against which success criterion will be solved?
2. Data and architecture. Sources, access, the RAG structure, model choice and privacy boundaries.
3. Small MVP. A measurable prototype working in a single flow.
4. Evaluation. Accuracy, cost, latency and user feedback.
5. Production. Monitoring, error handling, human approval and continuous improvement.