AI Integration

End any week, with one message. No notice period, no exit interview, no fine print. We invoice weekly, so the most you’re ever committed to is the current week.

It’s in your contract: tell us, and we refund that week. No questions, no invoices to dispute, no calls to escalate. The only rule: refunds apply to the most recent week.

Because hours are the wrong metric. If we’re optimizing for hours billed, we’re not optimizing for your outcome. The deal is simpler: every week, we earn the next one. If we don’t, you don’t pay. We’re free to spend zero hours or sixty. What matters is whether you’re blown away.

We work with operators, not lottery winners. If a request would require breaking physics, the law, or a third party’s systems, we say so, and if we can’t align, we walk. The guarantee is mutual: you can fire us any week; we can also fire ourselves.

Both, depending on what’s right. For 90% of business use cases, well-prompted frontier models (OpenAI, Anthropic, open-weights like Llama) outperform a custom fine-tune at a fraction of the cost. We reach for fine-tuning only when the task is narrow, the data is proprietary, and the cost math works. We’ll tell you which case you’re in honestly.

Three layers: structured outputs with JSON schema validation, retrieval-augmented generation grounding the model in your sources, and evaluation harnesses that score real responses against expected behavior on every deploy. We don’t ship AI features without a way to measure when they’re wrong.

Your data sits where you tell it to, and the products we build for you run under your own license with the AI provider, so the data-privacy terms are whatever you’ve signed. If you have an enterprise contract with OpenAI, Anthropic, Azure, etc., your data is contractually excluded from training. If you’re on a default tier, review the provider’s terms before you wire production data through it. For sensitive cases, we deploy open-weights models in your own cloud (AWS Bedrock, GCP Vertex, self-hosted) so the question never arises. We never use your data to train anything for anyone else.

Prototype: a week. Production-ready with evals, guardrails, and observability: 4–8 weeks. The slow part isn’t the AI, it’s everything around it: auth, billing, rate limiting, content moderation, audit logs. We’ve shipped enough to know where the time actually goes.

Depends on the build. For RAG and agents: LangChain, LangGraph, LlamaIndex, and the Vercel AI SDK on the frontend. For self-hosted inference: vLLM, Ollama, llama.cpp, Hugging Face Transformers. For evaluation: Braintrust, Phoenix, OpenAI evals. For observability: LangSmith, Helicone, Langfuse. We pick boring, proven tools over the hype cycle, the AI stack shifts every six weeks so we choose what we can rip out cleanly.

Prompt-engineering integration on top of an existing app: €5-15k. A RAG system over your own docs with evals and a real UI: €15-40k. A multi-step agent with tools, memory, and guardrails: €40-100k+. Runtime API costs are separate and depend on model and token volume. We budget the API spend into the proposal so you don’t get a six-figure surprise from OpenAI in month two.

When the same job is doable with a SQL query, a rules engine, or a form. When you need sub-200ms latency. When 100% deterministic output is non-negotiable (legal contracts, financial postings, medical dosing). When there’s no feedback loop to catch when the model is wrong. We’ll tell you ‘don’t do this’ if the use case doesn’t justify it, even if it shrinks the engagement.

Real agents. We build AI agents that call tools (function calling, MCP), execute multi-step plans (LangGraph state machines), read and write to your databases and APIs, and run unattended on schedules. Shipped examples: invoice-triage bots, internal-research agents, content-ops pipelines, automated QA harnesses. Chatbots are the boring case. Agents that move work forward are where the leverage lives.