Software
AI triage that reads the ticket before a human does
A two-person support team was spending most of its day categorising and routing tickets rather than answering the hard ones.
First-response time cut by more than half
AI App Development
AI features that survive contact with real users.
We build retrieval, agents and automation into products that already have customers, with the evaluations, guardrails and cost controls that separate a production feature from a conference demo.
AI App Development
The demo is the easy ten percent.
Anyone can wire a model to a text box in an afternoon. The work starts when the feature has to be right on the thousandth request, stay inside a budget, and fail in a way your support team can explain.
We treat an AI feature like any other piece of production software: it has acceptance criteria, a test suite, monitoring and an owner. The model is one dependency among several, and it is swappable.
Included
- RAG and search over your documents, tickets, catalogue or codebase
- Agents and workflow automation wired into the tools your team already uses
- Evaluation harnesses, prompt regression tests and human-in-the-loop review
- Model routing and caching tuned to keep unit economics sane at scale
What you end up with
An AI feature in production with evaluations, guardrails and a cost per request you can actually forecast.
The difference
Prototype vs. production AI
Most AI projects stall between these two columns. Closing that gap is the service.
AI prototype
- Works on the five examples it was built against
- Prompt edits shipped on instinct, regressions found by users
- No idea what a request costs until the invoice arrives
- One model, hard-coded, with no fallback
- Answers with confidence whether or not it has the facts
Production AI feature
- Scored against an evaluation set that grows with every bug
- Every prompt change runs through regression tests in CI
- Per-request cost and latency tracked and budgeted
- Routing, caching and fallbacks across providers
- Grounded in your data, cites sources, and knows when to hand off
What we build
What we build with AI
Four shapes cover most of what teams actually need. Many products end up combining two.
Retrieval and knowledge search
Question answering over your documents, help centre, tickets or codebase, with chunking, hybrid search and citations so answers can be checked.
Agents and workflow automation
Multi-step agents that read, decide and act inside your CRM, helpdesk or back office, with approval gates wherever an action is hard to undo.
Document and data extraction
Structured data out of invoices, contracts, forms and emails, validated against a schema and routed to a person when confidence is low.
AI features inside existing products
Drafting, summarising, classifying and recommending, added to the app you already have, behind a feature flag, measured against the workflow it replaces.
Process
How the work runs, phase by phase.
Each phase ends with something you can see, use or sign off.
Frame the problem
We pin down the task, the baseline it has to beat and what a wrong answer costs. Some problems turn out to need a rule and a database query, and we will say so.
Build the evaluation set
Real examples with known-good outputs, collected before any serious prompt work. It becomes the definition of done and the regression suite.
Prototype against the evals
Retrieval strategy, prompts, tools and model choice are iterated against scores rather than impressions, so progress is measurable.
Harden for production
Guardrails, rate limits, caching, fallbacks, tracing and cost budgets. Human review is added where the stakes call for it.
Launch and tune
A staged rollout behind a flag, with live traces feeding new cases back into the evaluation set. Your team gets the runbook and the dashboards.
Stack
The tools we reach for.
Defaults, not dogma. If you already run something else, we work within it.
- Models
- Claude
- OpenAI
- Gemini
- Open-weight models
- Orchestration
- LangGraph
- Claude Agent SDK
- MCP
- Temporal
- Retrieval
- pgvector
- Pinecone
- Elasticsearch
- Hybrid search
- Backend
- Python
- FastAPI
- Node.js
- PostgreSQL
- Quality
- Eval harnesses
- Prompt regression tests
- Tracing
- Cost dashboards
Timelines
How long it usually takes.
Typical ranges for well-scoped work. Discovery comes first and produces a fixed quote and a date.
- Feasibility spike
1–2 weeks
A scored prototype on your own data and a go / no-go recommendation.
- Single production feature
4–8 weeks
One workflow, evaluated, monitored and rolled out behind a flag.
- Multi-agent system
8–16 weeks
Several tools and approval paths, integrated with internal systems.
Case studies
Where we have done this before.
Industries
Sectors where this comes up most.
SaaS & Startups
From the first prototype to a platform that holds up under growth: a senior team for founders without one yet.
Healthcare
Telemedicine, patient portals and clinical tools built around real workflows and a privacy-first architecture.
Fintech
Payments, lending and finance tooling where correctness, auditability and security are the product.
FAQ
AI App Development: common questions.
Do we need to train our own model?
Almost never. Retrieval over your data plus a strong general model covers the large majority of business use cases, costs far less and improves every time the underlying model does. We recommend fine-tuning only when the evaluations show a gap that nothing else closes.
How do you stop it making things up?
You cannot remove the risk entirely, so we design around it: answers are grounded in retrieved sources and cite them, outputs are validated against schemas, low-confidence cases are routed to a person, and the evaluation set measures the error rate so it is a known number rather than a surprise.
What happens to our data?
It stays in your accounts. We use provider configurations that do not train on your inputs, keep retrieval indexes in your own infrastructure and document exactly what leaves your network and why.
What will it cost to run?
We measure cost per request during the prototype and give you a forecast at your expected volume before you commit to a build. Caching and routing cheaper models to easier requests usually bring it down substantially.
Can you add AI to our existing product?
Yes, that is the most common engagement. We work inside your repository and ship the feature behind a flag, so nothing changes for users until the numbers say it should.
Next step
Have something that needs building?
Tell us the problem in a few sentences. We will reply within one business day with an honest read on scope, cost and whether we are the right team for it.