Scoped against a decision, not a demo
We start from the decision the system must make and the cost of getting it wrong. That determines the accuracy bar, the architecture and the budget - before anyone writes a prompt.
Engineering
When nothing off the shelf understands your business
Some problems are specific to you: your terminology, your rules, your data, your edge cases. We build those systems from scratch - properly evaluated, integrated with what you run, and handed over with the source and the documentation so you are never locked in.
From$20,000
A bespoke system in your environment with an evaluation suite, measured baseline and full handover.
Indicative starting price. The fixed fee for your scope is quoted after the two-day diagnosis, before any build begins.
What makes it work
Three things we insist on
Evaluated before it is trusted
Every build ships with a held-out evaluation set drawn from your real cases, and a measured baseline. You know what the system gets right, what it gets wrong, and how that changes with each release.
Built to be handed over
Source, prompts, training data, evaluation sets and runbooks are yours. We would rather be retained because the work is good than because you cannot operate without us.
Scope one workflow, end to end An input, a defined done, a defined escalation, and a number that moves. Not a capability
Decide what it may do Drafts a reply or sends it. Recommends a refund or issues it. This drives the architecture and the audit logging you will wish you had
Build against your systems Real APIs, real permissions, structured output validated strictly against your own schema
Hand over everything Source, prompts, evaluation sets, runbooks. What you own is what you can change without us
Capabilities
What is actually included
- 01
Retrieval over your knowledge
Search and question-answering across your documents, tickets, code and records, with citations and access control preserved.
- 02
Fine-tuning and open-weight deployment
Smaller specialised models where volume, latency, cost or data residency rule out a hosted API.
- 03
Multi-step agents with real tools
Systems that plan, call your APIs, check their own work and stop when they are unsure.
- 04
Classical ML where it wins
Forecasting, classification, ranking and anomaly detection - often cheaper and more accurate than a language model for the same job.
- 05
Production engineering
Versioning, observability, cost controls, rollback and load testing. The unglamorous half that decides whether it survives contact with users.
In detail
What this covers, specifically
A category name is not a scope. These are the individual pieces of work inside this practice - take the two that apply to you and ignore the rest.
- 01
Retrieval-augmented generation (RAG)
Answering from your own documents and records by retrieving the relevant passages at question time and citing them, rather than relying on what a model memorised.
- 02
Fine-tuning
Adapting a smaller model to your domain, terminology and output format using LoRA or similar, where prompting alone cannot reach the accuracy bar.
- 03
Open-weight deployment
Running models inside your own cloud or on-premise, for cases where data residency, contractual terms or volume economics rule out a hosted API.
- 04
Multi-step agents
Systems that plan a task, call your APIs to carry it out, check their own work and stop when uncertain, rather than answering in one shot.
- 05
Classical machine learning
Forecasting, classification, ranking and anomaly detection using conventional models, which are frequently cheaper, faster and more accurate than a language model for the same job.
- 06
Structured extraction
Turning free text, email, documents or transcripts into typed, validated records that another system can consume.
- 07
Evaluation harness
A held-out test set built from your real cases plus the scoring code, so every future change can be measured instead of eyeballed.
- 08
Model routing and fallback
Sending each request to the cheapest model that handles it well, with defined behaviour when a provider is slow, rate-limited or down.
- 09
Integration middleware
The API layer, queues, retries and idempotency that sit between the model and your systems and decide whether it survives production.
What you receive
Concrete artefacts, not a slide deck
- Technical design with architecture and cost model
- Working system in your environment
- Evaluation suite and measured baseline
- Full source, prompts and documentation
- Handover and enablement for your team
01 Diagnose Days 1-2
02 Prove Week 1
03 Integrate Weeks 2-3
04 Operate Ongoing
Around three weeks end to end. That comes from scoping tightly to one workflow - not from skipping a phase. Each still ends in evidence you can check.
Further reading
Written up in more detail
- Multi-agent systems: when the complexity pays The three conditions that justify splitting one agent into several, the costs nobody budgets, and why most multi-agent designs should be a workflow instead.
- Structured output: getting reliable JSON from a model Schema-constrained decoding, permissive parsing with strict validation, and why a valid object with wrong values is the failure that actually costs you.
- Tool-calling agents: what to let them touch How to tier the actions an agent may take, why the dangerous combination is read plus write plus untrusted input, and the controls that survive a bad plan.
Let's talk
Bring us the problem, not the spec
The most useful first conversation is about the decision you need made better and what it currently costs you to get it wrong. We will tell you whether AI is genuinely the right tool - and what we would build if it is.
Related practices