Skip to content
Book a free 30-minute call with an engineer

AI & Machine Learning Development

AI features that do a specific job inside your product.

Chat Portal AI screens

The AI features worth building are the boring ones: answering support questions from your own docs, pulling fields out of invoices, drafting replies a person then checks, or finding the right record in a large archive. We look for the tasks in your product where AI saves real time and build those.

We work with OpenAI, Anthropic and open-source models, using Python, LangChain and vector databases. Every feature gets an evaluation set so we can measure accuracy, plus cost limits and logging, so you know what it does and what it costs before it goes to every user.

What’s included

Pick the parts you need. Most projects use all of them.

  1. 01

    AI assistants & chatbots

    Assistants that answer from your knowledge base and hand off to a person when unsure.

  2. 02

    RAG & semantic search

    Search over documents, tickets and files that understands meaning, with sources shown.

  3. 03

    Document automation

    Extracting data from invoices, contracts and forms into structured records.

  4. 04

    AI agents & workflows

    Multi-step automations that call your APIs, with approval steps where it matters.

  5. 05

    Predictive models

    Forecasting, scoring and recommendations trained on your own data.

  6. 06

    Evaluation & monitoring

    Test sets, accuracy tracking, cost dashboards and guardrails for production AI.

What you walk away with

  • Use-case review and a feasibility prototype
  • Production AI feature built into your product
  • Evaluation dataset and accuracy report
  • Cost controls, logging and usage dashboards
  • Documentation and a plan for improving the model over time
Chat Portal AI screens

Chat Portal AI, Conversation intelligence SaaS

Tools we use

We choose per project. These come up most often for ai development work.

summarise.tsTypeScript
1const res = await openai.chat.completions.create({​
2 model: "gpt-4.1-mini",​
3 response_format: { type: "json_object" },​
4 messages: [​
5 {​
6 role: "system",​
7 content: "Summarise support tickets. Reply as JSON: " +​
8 "{ summary: string, priority: 'low'|'high' }",​
9 },​
10 { role: "user", content: ticket.body },​
11 ],​
12});​
13​
14const result = TicketSummary.parse(​
15 JSON.parse(res.choices[0].message.content ?? "{}"),​
16);​

How a project runs

Five stages, two-week sprints and a demo at the end of every one.

  1. 01  Week 1

    Discover

    A couple of workshops with your team. We map the users, the must-haves and the risks, then send a written scope with a fixed estimate.

    MVP roadmapSprint 4 of 6 · launch in 3 weeksLater2Next2Now2Done3Referral codesGrowthDark modeUIAdmin exportsBackendPush remindersMobileStripe checkoutPaymentsOnboarding flowDesignSign up & loginAuthBooking APIBackendStore listingReleasev1.0 live
  2. 02  Weeks 2 to 4

    Design

    User flows, wireframes and the visual design in Figma. You click through a prototype, and we test it with a few real users before any code.

    Home · v2StaylineLisbon, 3 nights12 to 15 May · 2 guestsFree cancellation until FridayTotal€486Book nowColourTypeAaHeading 32Body 16 / 1.5Caption 12ComponentsPrimarySecondaryToggleTested with 6 users ✓
  3. 03  2-week sprints

    Build

    Code, tests and a staging site you can open any time. Every sprint ends with a demo, so you always see what changed.

    app.yourcompany.com/ordersNorthdeskOverviewOrdersCustomersInventoryReportsSettingsOrders142 this month · $48,210+ New orderOrderCustomerStatusTotal#1042Maple StudioNew$3,400#1038Harlow LtdPaid$1,240#1039BrightwayPending$860#1040Okafor & CoPaid$2,115#1041NordlineShipped$540Order #1042 created
  4. 04  Go-live

    Launch

    Security checks, load testing, App Store and Play Store submission, then a production release with monitoring switched on.

    ~/app — zsh$ git push origin mainEnumerating objects: 14, done.To github.com:acme/app.git 4f2a1c9..b81e0d3 main → mainCI triggered: pipeline #482Instances3 / 3 healthyCPU 64% · autoscale onPipeline #482main · b81e0d3BuildTestDeploy
  5. 05  After launch

    Improve

    We watch the analytics, fix what users trip over and keep shipping features. Most clients keep us on after launch.

    Uptime · last 30 days99.98%30 days agoTodayResponse time142 msp95 · all regions09:0015:00Incident resolved · 4 minCheckout API latencyRolled back · on-call: SamDeployed to productionv2.14.1Today 09:12Latest

Recent projects

AI Development: common questions

How much does it cost to add AI to our product?

A focused AI feature, such as a support assistant over your docs, usually costs $15k-$50k to build. Larger agent-based workflows or custom models cost more. Running costs depend on usage, and we estimate them up front.

Is our data safe with AI models?

We use providers and settings that don't train on your data, keep sensitive information out of prompts where possible, and can run open-source models in your own cloud when required.

How do you stop the AI from making things up?

We ground answers in your own content with retrieval, show sources, test against an evaluation set, and set the assistant to say it doesn't know rather than guess.

Which models do you use?

We pick per task. GPT and Claude models for reasoning and writing, smaller or open-source models for high-volume or private workloads. We often route between several to balance quality and cost.

Can AI work with our existing systems?

Yes. Most of our AI work involves connecting models to existing databases, CRMs, help desks and internal APIs.

Related services