03.1
WhatsApp & chat agents
Agents that complete a real task end to end — booking, qualifying, answering — on WhatsApp or web chat, with tool access to the systems that hold the answers.
Read moreAgents in production,
not prototypes.
Conversational agents on the channels your customers already use, document extraction, and automation between systems that were never designed to talk to each other.
An AI demo is easy to build and hard to trust. The distance between a convincing prototype and something you can leave unattended with customers is made up of the unglamorous parts: what happens at low confidence, what happens when the model is wrong, what it costs per conversation, and how you would even know if quality dropped.
So every build here ships with an evaluation harness, defined fallback behaviour and a human handoff path. If quality cannot be measured, it cannot be maintained, and an agent nobody is measuring is a liability pointed at your customers.
This is work we operate rather than only deliver. Mesa runs an unattended booking agent on WhatsApp and Kitchen Stocker runs an OCR pipeline over supplier paperwork, both in production, both ours.
Included
- Prompt, tool and fallback design
- Human handoff paths for low confidence
- Evaluation harness so quality is measurable
- Channel integration and rate limiting
- Cost monitoring per conversation
Not included
- Model inference costs, billed at provider rate
- Training bespoke foundation models
03.2
Invoice & document OCR
Invoices, delivery notes and forms read into records you can reconcile, with confidence scoring and a review queue for the ones the pipeline is not sure about.
Read more
03.3
Workflow automation
Moving data and triggering work across tools with no native integration — with retries, idempotency and alerting when a step fails.
Read more
03.4
LLM integration & evaluation
Model, prompt and tool design, plus the evaluation set that tells you whether a change improved things or quietly broke something else.
Read more
Proof
Mesa
An AI concierge handling restaurant bookings on WhatsApp, unattended, 24/7.
Read the case studyFrequently asked questions
Which models do you use?
Whichever fits the task, the latency budget and the cost ceiling — the choice is made at specification and is deliberately replaceable. We build behind a provider-agnostic layer so switching models is a configuration change rather than a rewrite, because this market moves faster than any build cycle.
What happens when the agent gets something wrong?
It should hand off rather than improvise. Every agent we build has a confidence threshold below which a human takes over, and the handoff is designed as a normal path rather than an error state. Conversations are logged so failures become evaluation cases instead of anecdotes.
What does this cost to run?
Inference is billed at provider rate and passed through without markup, and we instrument cost per conversation so the figure is visible from the first week rather than arriving as a surprise on an invoice.
Is our customer data used to train models?
No. We use providers and configurations that exclude your data from training, and the data handling is written into the specification. This is an EU business with genuine GDPR obligations, and vague answers here are not adequate.
Other services
01
Web design & development
Marketing sites, landing pages and web applications — built static or server-rendered depending on what the content actually needs.
02
Custom software & SaaS
Schema design through subscription handling, roles, onboarding and the admin surfaces that keep a product operable a year after launch.
04
Search & AI visibility
Ranking in a results page and being cited in an AI answer are now two different jobs with two different requirements. Most sites are built for neither.
05
UGC & creator content
Short-form video made by real creators, generated with AI where that is honest, and edited into the variants a paid social account actually needs.