The packages cover the eleven things we get asked for most. This lane is for everything else: the system nobody has built for you yet, the team that needs a senior pair of hands for three months, the architecture decision you can’t afford to get wrong. Same discipline — written scope, fixed quote, weekly demos — with the shape set by your problem instead of a catalogue.
Most AI projects fail on the engineering around the model, not the model. We do both halves — which is why the demo and the production system are the same system here.
Multi-step agents, tool use over your own APIs, orchestration and sub-agents, memory, background workers, and human approval wherever a mistake would cost money.
RAG that survives real documents, structured extraction from messy inputs, knowledge graphs where relationships matter, and citations so a human can verify in seconds.
Eval suites in CI, tracing across the whole loop, prompt-injection defence, permission policies, and the caching and routing work that cuts cost without cutting quality.
Web apps, dashboards, internal tools and APIs — auth, roles, billing, admin, background jobs. The unglamorous 80% that decides whether the AI part ever reaches a user.
Pipelines, warehouses, semantic layers, and integrations into the systems you already run — CRMs, helpdesks, accounting, messaging, and the internal API nobody documented.
Infrastructure as code, CI/CD, environments, observability, cost control, and the on-call story — in your cloud account, under your org’s rules.
Pick by how much certainty you have. Less certainty, shorter commitment — that’s the whole rule. Most clients start with a sprint and move to a retainer once the first thing is live.
A senior engineer on call for decisions: architecture reviews, build-vs-buy, vendor selection, hiring, code review, and the “is this a terrible idea?” question before you spend three months on it.
Two to four weeks, one written outcome, fixed price. We scope it together, build it in your repo, demo weekly, and hand it over working. The default way custom work starts here.
Two or three days a week inside your team — your standups, your board, your repo. For companies who need senior AI engineering now and are still hiring for it.
For a longer programme: we embed with your team and ship a system together over a quarter or more, with a written plan to make ourselves unnecessary by the end of it.
Custom doesn’t have to mean vague. You describe the problem once; we come back with a written scope, a fixed price for the first stage, and the assumptions it rests on. If those assumptions turn out wrong mid-build, we requote before doing the work — never after.
A call, an email, or the scope matcher — whatever’s quickest. Rough is fine; we ask the sharpening questions.
Within 48 hours: outcome, deliverables, stages, price for stage one, assumptions, and what’s out of scope. One page.
The first stage is always the riskiest unknown, not the easiest piece. You find out early whether this works.
Continue, pause, or take it in-house with the docs we wrote. No stage locks you into the next one.
Saying this up front saves everyone a call. If your project is on this list, we’ll point you somewhere better rather than take the work.
Fine-tuning and adapting open models for a specific task, yes. Pre-training your own frontier model is somebody else’s business and a much larger budget than ours.
The audit exists because a plan is useful. A hundred-page transformation deck with no working system behind it isn’t something we’d be proud to invoice for.
Scraped personal data, content you don’t own, or a model trained on a competitor’s material. We’ll help you find a legitimate data path instead.
A surprising share of “AI projects” are a missing integration, an unowned process, or a report nobody built. We’ll tell you that on the first call, even though it’s the version where we sell you less.
Send a paragraph about what you’re trying to build. You’ll get a real reply from the engineer who’d do the work — within one business day, with the questions that actually matter.