Assistants on your data
Chat and search that answer from your documents, help centre or database, with sources cited and permissions respected (RAG).
AI that earns its place in the product.
Assistants grounded in your own data, document and image understanding, and automation that removes real work. We pick the model, control cost and latency, and ship an interface people trust. If a simpler solution beats AI, we say so.
Model-agnosticClaude, GPT or on-device, chosen on your data
MeasuredEvaluations before launch, not vibes
In productionAI already ships inside our own apps
What we build
What this covers.
Document understanding
Read invoices, forms, contracts and PDFs; pull out the fields you need and route them where they belong.
Image & vision
Recognise text, objects and colours in photos and screenshots, tag them, and find similar images.
Workflow automation
Classify tickets, draft replies, summarise calls and move data between tools, with a human in the loop where it matters.
On-device, privacy-first AI
Apple Intelligence, Vision and on-device models when data must never leave the phone.
Evals, cost & guardrails
Test sets built from your real inputs, latency and cost budgets, monitoring, and fallbacks when a model has a bad day.
Proof
Already running in our own apps.
Lume turns photographed notes into study flashcards. Haberdashery reads text in screenshots, suggests tags and writes notes entirely on the device.
Lume — AI Flashcard Maker
Snap your notes, get smart flashcards. Study the smart way.
Haberdashery: Bookmark Library
A lovely library for links, photos, videos and notes — with on-device search and offline saves.
How we work
From brief to live.
01
Audit
We look at the task, the data and what “good” means, and tell you honestly whether AI is the right tool.
02
Prototype & eval
A working prototype against a test set built from your real inputs, so quality is measured, not guessed.
03
Integrate
Wire it into your product or back office with proper prompts, retrieval, permissions and error handling.
04
Monitor & tune
Track quality, cost and latency in production and keep improving as models change.
Tools we use
- Anthropic Claude
- OpenAI
- Apple Intelligence
- Core ML & Vision
- Embeddings & vector search
- Python
- Node.js
- Swift
Questions
Frequently asked.
What kinds of AI solutions do you build?
Four kinds: assistants that answer from your own documents (RAG), document and image understanding, content generation inside a product, and back-office automation. If a simpler solution beats AI, we say so.
Can you add AI features to software we already have?
Yes, it is our most common request. We work inside your existing codebase and data; a short technical review leads to a scoped proposal.
Which AI models do you work with?
We are model-agnostic: usually Anthropic Claude or OpenAI GPT for language, and smaller or on-device models where privacy or cost demands it. We benchmark on your real data before choosing.
How do you keep AI answers accurate?
By grounding answers in your own sources, measuring them against a test set before launch, showing sources to users, and monitoring quality after release.
Other services
Often paired with.
Have a project in mind?
Send a short brief — what you want to build, roughly when, and any constraints. We reply with honest next steps, scope and timing.