Knowledge and document systems
Help users find information across approved documents. Build search, question answering and source references around the material your team needs to use.
AI engineering & agentic systems
AI engineering & agent development. Working with your data and your tools.
Add AI capabilities to your software or build a dedicated system for a specific business process. We define what the model should do, which tools it can use and how its output will be checked, then build and evaluate it against agreed examples.
01 What we build
Intelligence, put to work.
AI features and agents that use your business information, connect to your tools and carry out defined tasks with the right checks and approvals.
Document processing
Extract, classify and structure information from documents. Validate the output and route exceptions for review before it enters your business systems.
Agents connected to business tools
Build agents that retrieve information, prepare updates and complete defined steps through your APIs and applications. Set permissions and approval rules for actions that change records or contact people.
Chat and voice agents
Answer customer enquiries by chat, DM or phone using your own services, prices and FAQs. Book into your calendar, log the conversation and hand over to a person when that is the right call.
Automated workflows
Connect email, calendars, CRMs and internal systems so routine steps run on their own. The model handles the language; the workflow handles the rules, retries and records.
AI inside your product
Bring AI search, assistance or analysis into your application. Design the user experience, evaluation criteria and fallback behavior together.
Model integration, tool connections, evaluation and application logic built around an agreed use case.
02 How we build it
AI you can depend on.
The model handles language. Code handles numbers.
Prices, availability, calculations and bookings come from your data through deterministic code, so the parts that cost money when they go wrong cannot be made up.
Answers cite what was retrieved
Knowledge systems return the sources they used, and the backend checks those citations against what was actually retrieved before showing them.
Evaluated before rollout
Representative examples and acceptance criteria are agreed up front. One document AI system is re-tested against a 95-question evaluation after every change; one booking agent runs 57 automated regression checks.
People approve what matters
Actions that change records, send messages or commit money can require a person's approval, and agents hand over to your team with the full context.
Models matched to each task
Stronger models where the reasoning is hard, faster and cheaper ones for scripted steps. Routing like this brought one voice agent to about $0.27 per call.
Hosted or private
Use hosted models through their APIs, or run open-source models on infrastructure you control when your data handling requires it.
Technologies in this work
- Claude
- Gemini
- Qwen3
- LangGraph
- n8n
- RAG
- pgvector
- OCR
- ElevenLabs
- Twilio
- PostgreSQL
- Python
03 Related work
Built, shipped and measured.
The requirements, technical decisions and results behind systems like this one.
04 Our approach
A clear plan. Working software. A documented handover.
01
Understand the requirements
Discuss users, business problems, existing systems and constraints. Define what a useful first release needs to accomplish.
02
Design and scope
Agree the architecture, deliverables, milestones and project price. Review deployment choices, data handling and running costs.
03
Build and evaluate
Review working software as development progresses. Test application behavior and evaluate AI outputs against agreed acceptance criteria.
04
Deploy and hand over
Launch in the agreed environment. Receive the custom source code, documentation and a walkthrough, with support responsibilities defined.
05 Good questions
Questions, answered.
Software that uses an AI model to choose and carry out steps toward a defined task, using approved tools and information. An agent might retrieve a record, review related documents and prepare an update for approval. We define its permissions and checks around the use case.
We choose models per task. Recent builds use Claude and Gemini models through their APIs, and open-source models such as Qwen3 running locally through Ollama and llama.cpp. The choice is reviewed against quality, cost, speed and where your data may be processed.
Yes. We review your codebase, data and APIs, then scope the feature, whether that is search, an assistant, document processing or an agent, with the user experience, evaluation criteria and fallback behavior defined together.
We agree representative examples and evaluation criteria for the task. We test outputs, tool use and failure cases, and define when the system should ask for clarification or human review. AI can make mistakes, so those behaviors are part of the design.
Running costs depend on the models, usage volume and hosting. We estimate them during scoping and list them separately from the project price. Routing simple steps to lower-cost models is one of the ways we keep them down.
07 Your next chapter
What do you want to build?
Tell us about your application, the system your team needs or the AI capability you want to develop.
A free 30-minute conversation about your requirements and a practical first release.