Copilots and assistants
An assistant for support agents, store teams, sales or internal ops that answers from your own documentation and can act inside your systems - with a clear record of what it did.
AI & GenAI practice
AI that survives contact with production
Most pilots stall because they were never wired into a system of record. We build the retrieval, the guardrails and the integrations that turn a promising demo into something your team opens on a Tuesday morning.
What we build
Six things clients actually ask us for
Not a research programme. Working software that removes a specific piece of manual effort or answers a question people currently guess at.
Retrieval over your knowledge
RAG done properly: chunking that respects document structure, hybrid search, re-ranking, permission-aware retrieval, and answers that cite the source paragraph.
Agentic automation
Multi-step workflows where the model plans and calls tools - order lookups, ticket triage, data entry - with approval gates on anything consequential.
Document and email processing
Extraction and classification for invoices, purchase orders, contracts and inbound mail, with confidence thresholds that route the uncertain cases to a human.
Forecasting and ML models
Demand, price and propensity models trained on your history, evaluated against the baseline you use today, and deployed where the decision is actually made.
Model and platform integration
Choosing and wiring in the model layer - hosted or self-run - with an abstraction that lets you change vendor later without rewriting the application.
Domain advantage
AI for retail, from people who know the data model
Retail AI projects usually fail on data, not on models. Knowing what an item hierarchy, a promotion, a partial return or a store transfer actually looks like in the source system is most of the work.
- Demand forecasting at store, channel and SKU level, including new and short-life items
- Markdown and price optimisation with margin and sell-through constraints made explicit
- Personalisation and recommendations grounded in real availability, not just affinity
- Store and workforce analytics that connect traffic, conversion and labour
- Returns, discount and refund anomaly detection for loss prevention teams
- Catalogue enrichment: attributes, descriptions and images generated then reviewed
How we get there
From use case to something in daily use
Triage the use cases
We score candidates on value, data readiness and tolerance for error, then pick the one or two worth building. Several usually turn out to be a reporting problem, and we say so.
Check the data and the access
Where the content lives, who is allowed to see it, how fresh it is and what shape it is in. This step decides whether the rest is weeks or months.
Pilot with real users
A narrow, working slice in front of a handful of people who do the job, instrumented so we can see what they ask and where it falls short.
Evaluate before you widen
A test set drawn from real questions, scored for accuracy, grounding and refusal behaviour. Regression runs on every change, so improvements do not quietly break something else.
Productionise
Deployment, monitoring, cost per interaction, prompt and model versioning, fallback behaviour, and an audit trail of actions taken on a user's behalf.
Adoption and measurement
Training, feedback capture, and a number agreed at the start that tells you whether it worked. If it did not, we would rather find out in week six than year two.
Principles
What we hold ourselves to
Grounded, or it says so
Answers cite their source. When the system does not know, it says it does not know rather than producing something plausible.
A human on consequential actions
Anything that spends money, changes an order or contacts a customer passes through an approval step until the evidence says otherwise.
Your data stays where it should
Retention, residency and access rules are part of the design, not something retrofitted after security review.
Measured against a baseline
We compare against how the job is done today. An improvement that cannot be demonstrated is not an improvement.
No single-vendor lock-in
The model sits behind an interface. Swapping provider should be a configuration change and a re-run of the evaluation set.
Code and prompts you own
Everything we build is handed over in your repository, documented well enough for your team to carry it forward without us.
Toolchain
What we build with
- Python
- LLM APIs
- Open-weight models
- RAG pipelines
- Vector databases
- Hybrid search
- Agent frameworks
- Evaluation harnesses
- LangChain-style orchestration
- FastAPI
- Docker
- Kubernetes
- AWS
- Airflow
- MLflow
- Observability & tracing
Bring us the use case you are unsure about
We will tell you whether it is an AI problem, a data problem or a process problem - and roughly what each version would cost you.