AI · Service
AI System Design
The model is one component. We design the system around it: the data it may see, the tools it may call, what a person must approve, how answers are evaluated and what each request costs.
AnatomyThe model is one of seven parts. We design all seven; the dark section below walks through each.
Who is AI System Design for?
For teams that want AI inside a real workflow, with the controls a business needs.
What do you get?
- A system design covering data flow, retrieval, tools, permissions and approval points
- An evaluation set built from your real cases, and a harness to run it
- A working prototype on your data, measured against that set
- A production plan with latency and cost budgets per request
How do you know it's done?
- Quality is a score on your own test cases, not a demo impression
- Every model call is logged, attributable and within budget
- When the system is unsure, it hands off to a person
What are the parts of a production AI system?
The model is one component. These seven parts decide what it can see, what it may do and when a person signs off.
01
Sources & access
Defines the documents, databases and tools the system may read, scoped per user exactly as in your own apps.
Skip it and: The assistant answers from data this user was never allowed to see.
02
Retrieval
Finds the right passage, not just a similar one. Tuned and tested on questions your team really asks.
Skip it and: You get confident answers built on the wrong document.
03
Context
Decides what goes into the prompt, in what order and with which instructions. Versioned and reviewed like code.
Skip it and: Behaviour changes whenever someone edits a prompt in production.
04
Model
Chosen per task on quality, latency, cost and where your data may be processed. Swappable behind one interface.
Skip it and: You're locked to one vendor's pricing, limits and outages.
05
Guardrails
Checks inputs and outputs for scope, personal data, policy and format. Failures route to a safe fallback.
Skip it and: One unusual input, and the wrong output leaves the building.
06
Human approval
Anything irreversible (sending, paying, deleting, deciding for a customer) waits for a person.
Skip it and: An automation makes a decision nobody can undo.
07
Evals & monitoring
Every change runs against your test set. Production calls are traced with their cost and latency.
Skip it and: Quality drifts, and nobody notices until a customer does.
What we measure, on your cases
- Answer accuracy against a labelled set
- Groundedness: every claim traceable to a source
- Hand-off rate: how often it asks a person
- Latency at p95, per step
- Cost per request
How does an engagement run?
- 01
Brief. A few minutes with our intake. Your context arrives before the call does.
- 02
Intro call. Thirty minutes. We ask the hard questions and tell you honestly whether we're the right fit.
- 03
Written scope. Scope, exclusions, milestones and a fixed price per phase, in writing, before any invoice.
- 04
Architecture. Diagrams and decision records you review and approve before the build starts.
- 05
Build. Working software on staging, demoed on a fixed cadence. You see progress, not status reports.
- 06
Handover. Code, infrastructure, docs and runbooks in your accounts. Ongoing support only if you want it.
How is it priced?
By scope. After the call you get a written scope with exclusions, milestones and a fixed price per phase. Payment follows accepted milestones. The budget range you pick in the brief isn't a quote; it helps us suggest a first phase that fits.
Which technologies do you use?
Our usual stack for AI System Design. We work with yours when you have one.
- Python
- FastAPI
- LLM APIs
- Vector search
- PostgreSQL
- AWS
AI System Design: questions and answers
What does AI system design include?
The whole system around the model: the data it may see, retrieval, the tools it may call, what a person must approve, how answers are evaluated, and a latency and cost budget per request.
How do you measure whether an AI system works?
On your own cases. We build an evaluation set from real examples and track answer accuracy, groundedness, hand-off rate, latency at p95 and cost per request.
Which AI model do you use?
It's chosen per task on quality, latency, cost and where your data may be processed, and it sits behind one interface so it can be swapped without a rewrite.
Start a brief for AI System Design.
A few questions, about three minutes. Nothing is sent until you review it and press Send, and the first call is thirty minutes to check fit, both ways.