# LLM and AI product features

> LLM integration adds language-model features — summarisation, classification, extraction, drafting, semantic search, chat — to an existing product. DevRaxo builds them with evaluation tests, guardrails, fallbacks and cost controls, and chooses the model and provider that fit your accuracy, privacy and budget requirements.

- Area: [AI & automation](https://devraxo.com/services/ai/)
- Price: from £2,400. Model usage is the client’s, billed by the provider at cost.
- Timeline: 4–6 weeks
- Page: https://devraxo.com/services/ai/llm-integrations/

## This is for you if
- You run a software product and your customers are asking for AI features.
- You have a working prototype that is impressive but unreliable.
- You need to control cost, privacy and quality, not just ship a demo.

## The problem
- **Demos hide the failure cases.** A feature that works on ten handpicked examples fails in public on the eleventh.
- **Costs scale with usage.** Model calls cost money on every use. Without limits, success can be more expensive than failure.
- **Provider lock-in.** Building directly around one vendor’s quirks makes the next price change or outage a crisis.

## How it works
1. **Define success as tests.** Before building, we write the evaluation set: real inputs and what a good output looks like. It becomes your regression suite.
2. **Pick the model on evidence.** We compare candidate models on your tests for accuracy, speed and cost, and show you the trade-offs.
3. **Build guardrails and fallbacks.** Input and output checks, timeouts, retries and a graceful fallback so a failure never breaks the product.
4. **Ship with cost and quality monitoring.** Usage caps, cost dashboards and quality sampling so you see problems before customers do.

## What you get
- An evaluation test set
- Model comparison report
- Production integration with guardrails
- Cost caps and monitoring
- Provider-agnostic architecture
- Documentation and handover

## Questions
### Which models do you use?
We choose per task: the major hosted models from several providers, and open models where privacy or cost demands it. The choice is driven by your tests, not preference.

### How do you handle sensitive data?
We minimise what is sent, choose providers with suitable data terms, and can run models in your own cloud account when required.

### Can you improve an AI feature we already built?
Yes. Often the fastest win is adding an evaluation set and guardrails to what exists, then fixing what the tests expose.

### What does it cost to run?
We estimate cost per use from real tests and set a ceiling, so you know the monthly range before launch.