Home / Services / AI solutions
Agents, LLM features, document pipelines and workflow automation, scoped as a pilot on your own data, measured for accuracy and running cost, then rolled out. For finance, healthcare and public-sector clients the same work runs on self-hosted models with nothing leaving your environment.
The demo is easy. Someone wires a prompt to your data, it answers three questions convincingly, and the project gets approved. Then it meets real inputs: a scanned invoice at an angle, a customer who asks two things at once, a document in a format nobody anticipated. The accuracy nobody measured turns out to be somewhere around seventy percent.
Seventy percent is the worst possible number. Too good to abandon, not good enough to trust, so a person ends up checking every output and the automation has added a step instead of removing one.
The second thing that stalls projects is cost. Token spend is invisible during a pilot and very visible at ten thousand documents a month. We have seen a working pipeline shelved because nobody priced it before building it.
And for a large share of the companies that approach us, none of it matters until one question is answered: where does the data go? Banks, hospitals, insurers and government suppliers all start there. If the answer involves a third-party API, the conversation ends.
Sound familiar?
A proof of concept that impressed everyone and has been parked for six months.
Nobody can state the accuracy rate, because nobody built a test set.
Monthly API spend that nobody forecast and nobody can now explain.
Legal or compliance blocked it, and no one offered a version that clears review.
It works for the happy path and fails silently on everything else.
The vendor owns the prompts, the pipeline and the evaluation data.
Not a generic demo. We take a sample of your real documents, tickets or records, build against those, and show you where it fails as well as where it works.
Where the task can be scored objectively, we build a labelled evaluation set with your team. It stays with you, along with a way to re-run it whenever a model or prompt changes.
Token cost per document or per conversation, multiplied by your actual volume. If the economics do not work, we say so during the pilot, not after the invoice.
Rate limits, fallbacks when the model is unavailable, a confidence threshold that routes uncertain cases to a person, and logging you can audit.
Accuracy drifts as your inputs change and as providers update models. You get dashboards and alerts, not a silent degradation nobody notices for a quarter.
Open-weight models running inside your VPC or on your own hardware. Scoped case by case, since the right approach depends on your infrastructure and the model size, and it is the only version some clients can approve.
Every engagement starts with a paid pilot. It is the cheapest way to find out whether AI is the right tool for the problem you have, and sometimes the honest answer is no.
What exactly is being automated, and what accuracy makes it worth doing? A claims triage that must be 99% right is a different project from a draft-reply generator that can be 85%.
Output: a written task definition and a target accuracy.
Where the task allows it, we take a representative sample of your real data and label the correct answers with your team. This is the unglamorous part and it is why the rest works.
Output: a labelled test set you keep.
We build the pipeline, run it against the test set, and report accuracy, failure modes and cost per unit. You see the failures, not just the wins.
Output: a pilot report with go or no-go.
Production deployment with confidence thresholds, human review routing, fallbacks, logging and monitoring. Then a handover so your team can operate it.
Output: a running system and its documentation.
Model choice is a cost and privacy decision more than a capability one. Most tasks do not need the largest model available.
Claude, GPT and Gemini APIs
Default for anything where data can leave your environment. We benchmark two or three against your test set rather than assuming one is better.
Llama and Mistral, self-hosted
Open-weight models on your own infrastructure. Chosen when compliance requires it, not because they outperform the hosted ones.
Python, LangChain and LlamaIndex
Pipeline orchestration, retrieval and chunking. Kept thin. A wrapper you cannot debug is worse than no framework.
pgvector and Qdrant
Vector search for retrieval. pgvector when you already run Postgres and want one less system to operate.
n8n
Workflow automation connecting the model to the tools you already use: CRM, email, storage and ticketing, without a custom integration for each.
Langfuse
Tracing and evaluation in production. Every call logged with its inputs, cost and latency, so a drop in quality is visible.
Not fixed to these
This is what we reach for by default, not a limit. Most AI work has to fit around a system that already exists. If your stack is different, we work in it: a different cloud, a model your security team has already approved, an orchestration tool you are committed to. Tell us the constraints on the scoping call and we will say honestly whether they change the approach or just the tooling.
AI work is scoped in stages rather than as one project, so you can stop after the pilot if the numbers do not justify going further.
| Engagement | Typical timeline | Best for |
|---|---|---|
| Scoping call and written brief | 1 week | Deciding whether there is a project here at all |
| Pilot on your data | 2 – 4 weeks | One task, measured, with a go or no-go |
| Production rollout | 6 – 14 weeks | A pilot that cleared its accuracy bar |
| Self-hosted deployment | Scoped per project | Regulated data that cannot reach a third-party API |
| Monitoring retainer | Ongoing | Keeping accuracy and cost under watch after launch |
The pilot is priced on its own, so the decision to continue is made with a measured accuracy figure and a running-cost forecast in front of you rather than on an estimate. Model running costs are separate from build cost and depend on your volume.
Every project is priced in writing after a scoping call. You get a fixed number, with what is included and what is not, so there is nothing to negotiate later.
Tell us what you need built and we will send a scope and a price. The reply comes from the person who would run the project.
Get a quote →This page describes how we prefer to work, not a fixed contract. Deliverables, stages and timelines vary with the size and shape of the project, and some of what is described here will not apply to yours. The written scope you receive after the scoping call is the version that counts. How we engage →
With hosted models, to that provider's API (Anthropic, OpenAI or Google) under their enterprise terms, which exclude training on your inputs. With a self-hosted deployment, nowhere. The model runs inside your VPC or on your hardware and no request leaves your network. We will tell you which is realistic for your compliance position before quoting.
Where the task can be scored objectively, we build a labelled test set from your real data before writing the pipeline, agree a target with you, and report the measured figure against it. You keep the test set. Some tasks, open-ended drafting for instance, cannot be scored that way, and we will say so rather than inventing a number.
That depends on volume, and we forecast it during the pilot: cost per document or per conversation, multiplied by your real throughput. If the running cost makes the project uneconomic we tell you at the pilot stage.
Then we say so and you have spent a few thousand dollars instead of forty. It does happen, and often the problem turns out to need a rules engine, better data entry, or a fixed process rather than a model.
You do. Prompts, evaluation data, pipeline code and infrastructure configuration are handed over in your repository under your account. There is no licence, no hosted layer you have to keep paying us for, and no lock-in.
Usually yes. Most of our AI work sits alongside an existing system rather than replacing it: a classification step on incoming email, extraction from uploaded documents, a drafting assistant inside a tool your team already opens every day.
Every enquiry gets a reply within one working day, from the person who would run your project.