03 AI integration
AI features that survive contact with production.
Retrieval-augmented assistants, semantic search, agentic workflows, and LLM features wired into real products, with evals, guardrails, and cost discipline. We treat AI as an engineering problem, not a demo. If a feature should not use a model, we will tell you that too.
02 The position
The demo always works. Production is where AI features go to die.
It is trivial to wire a model to a prompt and produce a magical-looking demo. It is hard, and it is engineering, to make that feature reliable, affordable, and safe with real users, real data, and real edge cases. The gap between the demo and the product is where most AI initiatives quietly stall.
We build for the production side of that gap. That means retrieval grounded in your actual data, evaluations so you can measure whether a change made the feature better or worse, guardrails against the failure modes that embarrass you, and cost controls so a successful feature does not become a runaway bill.
It is the same standard we hold for any software or web application we ship. AI does not get a pass on rigor because it is new.
03 Philosophy
We are not here to put a chatbot on your homepage.
The market is awash in AI theatre, features added because the board asked for AI, not because they help anyone. We are openly skeptical of that, and our writing says so. AI is a genuinely powerful tool for the right problems, and an expensive liability for the wrong ones.
So we start with the problem, not the model. Is there a real task where retrieval, classification, generation, or an agent earns its keep? If yes, we build it properly. If no, we will say so and save you the budget, because a boutique studio's reputation is worth more than one more billable AI feature.
AI is a tool, not a strategy. The question is never 'can we add AI?' It is 'does this specific feature get better with it?'
04 Capabilities
What we build with AI.
- 01
Retrieval-augmented assistants (RAG)
Assistants grounded in your own documents and data, so answers are accurate and attributable instead of confidently wrong.
- 02
Semantic & hybrid search
Search that understands meaning, not just keywords, wired into your product with the relevance tuning that makes it actually useful.
- 03
Agentic workflows
Multi-step automations where a model plans and acts within hard guardrails, for the workflows that genuinely benefit from autonomy.
- 04
LLM features in-product
Summarisation, drafting, classification, and extraction built into the surfaces your users already use, engineered like any other web-app feature.
- 05
Evals & observability
The unglamorous core: test sets, scoring, and monitoring so you can prove a change helped, and catch regressions before users do.
- 06
Guardrails & cost control
Input/output safety, rate and budget limits, caching, and model-routing so the feature stays safe and the bill stays sane.
05 Method
How we think about the work.
- 01
Problem first, model last.
We will not add AI because it is fashionable. We identify a real task where a model measurably helps, and only then choose the smallest approach that solves it.
- 02
If you can't evaluate it, you can't ship it.
Every AI feature we build comes with a way to measure quality. Without evals, you are not engineering, you are guessing in production, and guessing is how AI features erode trust.
- 03
Guardrails and cost are part of the feature.
Safety and spend are not afterthoughts. We design the boundaries and the budget alongside the capability, because an unbounded AI feature is a liability waiting to surface.
07 Related thinking
Where this work goes deeper.
· Questions we get
Common questions, honest answers.
-
Do we need our own model, or do you use existing ones?
Almost always, the right answer is to build on existing frontier or open models rather than train your own, the cost and complexity of a custom model is rarely justified. We focus on the engineering around the model, retrieval, evals, guardrails, and integration, which is where the real value and difficulty live.
-
How do you keep AI costs under control?
Through model routing (using the smallest capable model per task), caching, prompt and context discipline, and hard budget limits. We design cost controls into the feature from the start, so a popular feature does not become a runaway expense.
-
What if AI isn't actually the right solution?
Then we will tell you, before you spend on it. A large share of 'AI problems' are better solved with conventional software, better search, or a clearer interface. We would rather lose an AI line-item than build you a feature that does not earn its keep.
-
Can you add AI to our existing product?
Yes, that is most of this work. We integrate AI features into existing software and web applications, wiring them to your real data and your real interface, with the evals and guardrails that make them safe to ship.
· Working together
Have an AI feature in mind? Let's pressure-test it.
Tell us the task you're trying to improve. We'll tell you honestly whether a model is the right tool, and if it is, how we'd build it to survive production.
Start a conversation