LLM app development built for production, not demos
We design and build applications on Claude, GPT, Gemini and open-weight models with routing, tools, retrieval, evals and observability in the architecture from day one.
Prototypes call a model directly from a web form. Production systems put four things between the user and the model: a gateway, an orchestration layer, a router and a measurement layer. Skip any of them and you will add it later under pressure.
Reference architecture. The router lets the same application use Claude for agentic work, a cheaper model for classification and an open-weight model where data must stay in-region. Everything emits traces to the measurement layer.
Customer and employee copilots that cite sources and know when to hand off. Built with our AI knowledge base practice.
AI features inside existing products
Summaries, smart search, generation and recommendations added to a SaaS product or mobile app without rewriting it. See AI integration.
Model choice
How we think about the four model families
Benchmarks change monthly; the qualitative differences change slowly. This is the thinking behind our model routing policy. The final decision for your application is made against your own eval set.
Model family
Tends to fit
Strengths we lean on
Watch-outs
Where it runs
Anthropic Claude
Agentic workflows, long documents, coding assistants, tool-heavy apps
Reliable tool use, strong instruction following, long context, MCP support in the ecosystem
Design prompt caching and context budgets early to keep costs predictable
Anthropic API, major cloud marketplaces
OpenAI GPT
General assistants, structured output, broad consumer-style chat
Mature SDKs and tooling, wide third-party support, multimodal input
Frequent model generations; pin versions and re-run evals on upgrade
OpenAI API, Azure
Google Gemini
Multimodal workloads, organisations already on Google Cloud or Workspace
Large context, native image, audio and video input, tight Google Cloud integration
Fewer community patterns for some agent frameworks
Google AI Studio, Vertex AI
Open-weight (Llama, Mistral)
Data residency, very high volume, custom fine-tuning, offline or edge use
Full control over data and weights, predictable unit cost at scale, no vendor lock-in
You own the operations: GPUs, quantisation, scaling, safety tuning and evals
Staged rollout behind a flag, runbook, handover and a production eval baseline.
What goes wrong when this is skipped
The three failures we are most often called in to fix: a prompt nobody can change because there is no test set, a bill that doubled because every request carries the whole document, and an assistant that follows instructions hidden inside the content it reads. All three are architecture decisions, made or not made in the first fortnight.
Usually more than one. We put a model router behind your application so each task runs on the model that suits it, and so you can switch when pricing or quality changes. The table on this page explains how we think about Claude, GPT, Gemini and open-weight models; the final choice is made with your eval set, not by opinion.
How long does an LLM application take to build?
A thin production slice typically takes four to eight weeks: one feature, real data, evals, security review and monitoring. Broader applications with several integrations usually run three to six months. We stage releases so something useful ships early.
Do you use frameworks like LangChain or build from scratch?
We prefer small, explicit orchestration code with the vendor SDKs, and add a framework only where it genuinely removes work. Frameworks hide the control flow you will need to debug at 2 a.m. Either way, you own the code.
What does an eval set look like and who writes it?
A versioned collection of real inputs with expected outputs or grading rubrics, scored automatically on every change. We draft the first set with your domain experts in week one and grow it from production traffic. See fine-tuning and evals.
Can you host on our own infrastructure?
Yes. We deploy to Cloudflare Workers, Azure, Docker on your VMs or on-premise, and we can run open-weight models where data residency rules require it. See our guide to on-prem versus cloud LLMs.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.