Service

AI feature integration for existing SaaS products

Your product already works. You need AI inside it — summarisation, semantic search, an assistant, automated generation — without rebuilding the thing that already works. I integrate LLMs into live SaaS products, handling the parts that actually make AI features hard: cost control, latency, streaming, evaluation, and graceful failure.

Start a conversation

typical time to ship an AI feature to production: Daystypical time to ship an AI feature to production
AI content platform built and scaled: Creatiz.aiAI content platform built and scaled
faster page load after a Creatiz.ai revamp: 40%faster page load after a Creatiz.ai revamp

Where teams get stuck

The demo worked, production did not

A prompt in a notebook is not a feature. Rate limits, token costs, streaming UX, retries, and hallucination handling are where AI projects stall — and they only appear once real users arrive.

AI costs are unpredictable

Without caching, model routing, and token budgeting, one power user can cost more than they pay you. I build cost controls in from the start rather than after the first alarming bill.

Your team has never shipped AI before

Choosing between fine-tuning, RAG, and prompt engineering is an architecture decision that is expensive to reverse. Most teams over-engineer it — usually RAG plus good prompts is enough.

What you get

LLM integration with OpenAI, Anthropic, or open models — with fallback routing
Retrieval-augmented generation over your own data, with a real vector store
Semantic and hybrid search that beats keyword matching on your content
Streaming responses with proper loading, cancellation, and error states
Token budgeting, caching, and per-user rate limits so costs stay bounded
Prompt versioning and an evaluation harness so quality does not silently regress
Monitoring on latency, cost per request, and failure rate

How it works

01
Feasibility review
I look at your product and data, and tell you honestly which AI feature is worth building first — and which ones will not survive contact with real users.
02
Thin vertical slice
One feature, end to end, in your real codebase. Working in production within days, so you learn from actual usage instead of a mockup.
03
Hardening
Cost controls, caching, evaluation, and failure handling. This is the part that separates a feature that ships from one that quietly gets rolled back.
04
Expansion
With one feature proven and instrumented, additional AI surfaces are fast to add on the same foundation.

Frequently asked questions

Can you add AI to a product you did not build?

Yes — most of this work is on existing codebases. I start by reading your code and data model, then ship a narrow feature end to end before touching anything broader.

Which AI models do you work with?

Claude, GPT, and open models via hosted inference. I usually route between them by task: a cheaper model for classification and extraction, a stronger one for generation and reasoning. That routing alone often cuts cost by more than half.

How do you keep AI costs under control?

Response caching, prompt compression, model routing by task complexity, and hard per-user token budgets. Cost per request is monitored from day one, not discovered on an invoice.

Do I need fine-tuning?

Almost certainly not. Retrieval over your own data plus well-structured prompts solves the large majority of product use cases at a fraction of the cost and complexity. I will tell you if you are the exception.

How long does an AI integration take?

A first production feature typically takes one to three weeks depending on how clean your data access is. Simple integrations — summarisation, classification, generation — often ship in days.

Want AI in your product without a rewrite?

Send me your product and the feature you have in mind. I will tell you if it is a week of work or a quarter — and which one is actually worth doing.

Email me Contact form

Replies within one business day · NDAs signed as standard