000 / Custom Apps
Custom AI Desktop App
A Windows app where someone presses a button, and a workflow runs that used to take five tools, three logins and ten minutes of focus. AI-assisted, installed locally, no cloud lock-in.
001 / Deliverables
What you get
What you end up holding, four concrete deliverables.
Installed app, working
A single .exe with everything bundled. No separate Python install, no PyPI adventures. Optional auto-update mechanism.
AI integration, debounceable
Whether Gemini, Claude, GPT or a local model, the app uses it as a tool, not as an end in itself. With caching, error handling, and sensible fallbacks if the model goes silent.
Hardware integration
Webcam calibration with ArUco markers, USB devices, printers, scanners, anything attached to a Windows machine. On the Book Lister project: webcam measures book dimensions automatically in under a second.
Tests + documentation
~250 pytest tests in the Book Lister codebase, because hardware + AI + live APIs go voodoo fast otherwise. Plus code documentation for you or your team if someone later needs to extend it.
002 / How it runs
How a custom-app project runs
-
01
Discovery
I look at the manual workflow to be replaced. Which steps, how much time, what goes wrong. Out of that comes the scope.
-
02
Prototype
A small vertical slice within one or two weeks, the most important path, end to end. You test, give feedback.
-
03
Build
On top of the prototype the app gets built out. Iterative, with intermediate builds you can test.
-
04
Handover
Installer + source + documentation. Plus a one-hour setup coaching for the end users.
003 / Price
Per project, in tiers
€8,500 is the entry for a focused app with one AI model, one hardware component and one API integration. On the Book Lister it was Gemini Vision + webcam + eBay APIs.
Larger apps with multiple workflows, multi-user architecture or cloud sync scale accordingly. The fixed figure comes after discovery, you know it before you sign.
Custom-app project
One workflow, one AI, one hardware component.
- Discovery + scope definition
- Vertical-slice prototype (2 weeks)
- Full app with installer (.exe)
- Pytest coverage for critical paths
- Setup coaching (1 h) + code docs
Reply within 24 h
004 / Reference
What such a project looks like
In early 2026 I built an app for a used-book seller. Before: 10 minutes per book, manually. After: 30 seconds.
30 seconds per book instead of 10 minutes
ArUco-calibrated webcam measures dimensions, Gemini 2.5 Vision reads title/author/publisher, Google Books + eBay Browse APIs set the price, eBay Trading API publishes the listing. ~6,450 lines of Python, 260+ pytest tests. 400 % throughput.
Read the full case study005 / Frequent questions
Before you inquire
Cloud AI or local model?
Depends on the workflow. For vision tasks, cloud (Gemini 2.5 / Claude) is usually better today. For sensitive data or offline requirements: local models (Llama-based, Ollama). We decide in discovery.
Who hosts the app?
Nobody, the app runs locally on the Windows PC. Cloud components only if explicitly wanted. No monthly server fees from me.
Who pays the API costs?
You, because the keys are on your account. On the Book Lister it was ~€5 per month for Gemini + Google Books. Scales with volume, but mostly manageable.
What if the AI hallucinates?
The question that comes up early in every discovery. Answer: pre-validation on the data side (e.g. ISBN check against Google Books), confidence scoring, and critical actions require user confirmation. AI as a suggestion, not as autopilot.
What breaks when the app glitches?
A crash handler writes a crash file per incident, the app remembers the last good state. On re-open the unfinished workflow is restored where possible. Data loss is extremely unlikely.
Can I extend the app myself?
Source comes with it, plus code docs. If you or your team know Python, yes. Otherwise we can also do a maintenance retainer, cancellable monthly.
006 / Related services
Often booked alongside this service:
Automation
Browser automation at production level
Browser automation that runs in parallel, fault-tolerant, without daily babysitting. Multi-worker with IPC, crash recovery, CDP-based auth.
Consultation
Discovery Call & Architecture Consultation
30 or 60 minutes on the phone, workflow, feasibility, architecture recommendation. Written summary. Price credited against any follow-up project.
A workflow that's too valuable to stay manual?
Briefly describe the workflow you want to automate, I'll be back within 24 h with an estimate and a proposed call time.
Send inquiry