Agents in production: permissions, guardrails and the approval step
Agentic AI is the most requested and least understood feature of the year. Here is how we ship agents that take real actions without taking the business with them.
Read postWriting
25 posts by iLytix Engineering.
From a broad AI idea to a first launch: how we plan an AI product
Turning a vague ambition into a product direction, a feature list, a roadmap and a first release that can actually ship.
Read post
Prompt injection is the new SQL injection: a threat model for LLM applications
The security questions we ask of every AI feature, the attack that nearly worked, and the controls that hold when the model does not.
Read post
RAG that answers correctly: what we changed after the first three systems
Retrieval-augmented generation is easy to demo and hard to trust. The fixes were in chunking, evaluation and citations, not in the model.
Read post
Evaluating an LLM application before launch, not after the complaints
How we build an evaluation set, what we measure, where LLM-as-judge helps and where it misleads, and the release gate that came out of it.
Read post
Controlling LLM costs without degrading the product
The bill that surprised a client, the four levers that brought it down, and why cost is a design property rather than a procurement problem.
Read post
Observability for AI systems: tracing, drift, cost and the feedback loop
You cannot debug what you cannot see. What we record for every AI request, how we watch for drift, and why the feedback loop is the point.
Read post
Designing for the day you change model provider
Providers change prices, models and terms without asking. The abstraction, evaluation set and routing layer that let us switch a production system in a week instead of a quarter.
Read post
Finding the workflows worth automating
How we identify the repetitive tasks, approval flows and document reviews that are ready for automation, and the ones that only look ready.
Read post
Document automation, from OCR to approval, with an audit trail that holds up
Invoices, claims and applications arrive as documents. Here is the pipeline we build to turn them into decisions, the step that kept going wrong, and the controls an auditor will ask about.
Read post
Where knowledge assistants actually earn their keep
Answering questions, searching documents and supporting customers sound similar. They are three different products with three different failure modes.
Read post
Prompting, retrieval or fine-tuning: how we choose
Three ways to make a model do your job, in order of cost. The project that asked for fine-tuning and needed a better prompt, and the one where fine-tuning was the only answer.
Read post
Start with clarity
Book a practical project discussion.
Share your product, AI, workflow, or data challenge. We will help identify the right starting point, delivery risks, scope, timeline, and next step.
Book a Project CallOr email sales@ilytixservices.com