AI engineering · London
Most AI demos answer questions. We build agents that change things — creating orders, filing shipments, resolving exceptions — behind approval gates your backend enforces, with the evaluation work to show they get it right.
What we actually build on
Model-agnostic by default
Vercel AI SDK
Anthropic
OpenAI
AI Gateway
MCP
Zod / Pydantic
Typed end to end
Next.js
React
TypeScript
tRPC
FastAPI
Tailwind
Retrieval and state
PostgreSQL
pgvector
Drizzle
Redis
Neon
Ship and observe
Vercel
AWS
Docker
Stripe
Playwright
Vitest
Also experienced with LangChain · LangGraph · LangSmith · Django · Celery · scikit-learn · ChromaDB
Case studies
Not pilots. Software a business depends on daily, written up in enough technical detail that you can judge the engineering.
Mailboxes Etc (Fortidia)
A conversational operations platform where branch staff run real shipping work through an AI agent — creating shipments, itemising customs declarations, and tracking parcels across carriers, with every write action gated in the backend.
Read case study
Mailboxes Etc (Fortidia)
Thousands of EPOS products mapped into a deep Square category hierarchy using embeddings for the easy cases, a frontier model only for the genuinely ambiguous ones, and a hard reject below a confidence floor.
Read case study
Atlantis Print
A print-on-demand commerce platform with a public ordering agent that places real orders and an internal credit-metered assistant — plus the evaluation pipeline that measures whether either is working.
Read case study
AI roles
Named after the job rather than the technology, because what matters is what the thing does and where it stops. Each one is running in production for a client today.
01
Takes the enquiry, quotes from your live prices, and places the order.
Not a chatbot that promises somebody will call back. It reads your real catalogue, handles the variants and quantities that make quoting slow, captures the customer, and puts the order through.
Where it stops. Every action is a typed tool with a validated schema, and write access is gated in your backend rather than requested in a prompt.
Live
Prices read from your catalogue at the moment of asking, never recalled from memory
02
Watches what is in flight, catches what has gone wrong, and writes the customer update before they ring you.
The work nobody has time for: re-checking every order across every system it touches, spotting the one that has stalled, and telling the customer first. It runs on a schedule rather than waiting to be asked.
Where it stops. It drafts the message. A person reads it and sends it.
Drafted
Customer updates written and queued for a person to send, not sent automatically
03
Does the filing nobody wants — classifying, coding, itemising — and hands it over for approval.
Customs declarations, tariff codes, catalogue mapping, quote triage. The tedious, high-volume, error-prone work that quietly eats skilled people, done in proportion to how hard each case actually is.
Where it stops. Below its confidence floor it refuses and asks for a human, rather than producing a confident wrong answer.
Thousands
Products mapped into a new hierarchy without anyone hand-sorting them
Services
We would rather be genuinely good at a short list than adequate at everything with an AI label on it.
Products
The hard parts of AI engineering — cost metering, provider fallback, retrieval quality — we solve on our own bill first.
KithFlow
LiveUK company intelligence, built for agents to consume
HyperSaaS
LiveDjango + Next.js SaaS foundation
KognitivAtlas
LiveAn AI travel planner that actually writes the plan
LLM-Wiki
Open sourceA knowledge base an agent operates
Engagements
Most clients begin with a fixed-price audit, so the first thing you buy is a decision rather than a commitment.
Fixed price
You know AI should be doing something in your business but not what, and you would rather not find out by spending six figures.
From £6k
You have identified one workflow worth automating and want it in production, not in a slide deck.
From £25k
You are building an AI product, or embedding AI deeply enough into an existing one that it needs its own engineering.
Next step
Most engagements start with a fixed-price audit, so the first thing you buy is a decision rather than a commitment.