CLIENT PROJECT
PropDesk — AI Staff Assistant
Staff at a property management company describe a tenant issue in plain words and get vendor + tenant drafts with data gaps flagged. Human reviews, nothing auto-sends.
Read the case →Home › Services › AI automation
// service
GPT and Claude doing real work inside your process
The useful version of "AI" in a small business is not a chatbot on the homepage — it is a model sitting inside an existing workflow: reading the email, scoring the lead, drafting the reply, writing the Monday report. I build those steps with a human review where the output goes to a customer, and with the model saying "I don't know" instead of inventing data.
from $300Describe your process →
What I build
- Classification and routing — inbound emails, form submissions or support messages sorted and sent to the right person or queue.
- Drafting with review — vendor requests, customer replies, quotes: the model drafts, a human edits and sends.
- Reports and summaries — weekly numbers from Sheets or your CRM turned into a readable report, delivered on schedule.
- Voice and content pipelines — voice note → transcript → post copy → image → published, with a check step.
- Internal assistants — a web app over your documents and records (see PropDesk) when a chat window is the right interface.
How I keep it honest
- Structured outputs and validation — the model fills fields, it doesn't free-write into your CRM.
- Missing data is flagged, not guessed (PropDesk reports what it doesn't know about a building).
- Human-in-the-loop for anything that leaves the company under your name.
- Costs estimated per run before the build; prompt caching and small models where they are enough.
- Your API keys, your accounts; the prompts are documented and handed over.
How a project runs
- Discovery call (free, 30 min). You walk me through the process; I ask about volumes, tools and who maintains the result.
- Written scope. Every step mapped, tools named, what "done" looks like, timeline and a fixed price. Nothing starts before you approve it.
- Build with test data. Error handling, retries and alerts are part of the build, not an extra.
- Review. You test against the scope checklist; fixes to anything off-scope are included.
- Handoff. Walkthrough video, step-by-step documentation, credentials in your accounts. Then 30 days of support.
Related work
3hsaved every week
CLIENT PROJECT
AI Report Generator
Google Sheets data → Claude API summary → formatted report in the owner's inbox every Monday at 9:00.
Read the case →
CLIENT PROJECT
AI Content Machine
Telegram message (photo, text or voice) → Whisper transcript → GPT-4o post copy → DALL·E image → Google Drive + Sheets → finished post back in Telegram.
Read the case →
OWN TOOL
Source on GitHub ↗Real-Time Interview Copilot
Windows desktop app: faster-whisper transcription → DeepL translation → Claude agent with a RAG knowledge base deciding when to answer, ask or stay silent.
Open demo →Questions
Our case is pretty unusual — can you still help?
Tell me the details on the discovery call. If no-code tools can't cover an edge case, I say so honestly and we figure out a workaround — sometimes a small custom script, sometimes a different tool. I don't take projects I can't actually deliver.
How much does this actually cost?
Fixed price, agreed before I write a single scenario — no hourly billing, no scope-creep charges. Simple integrations start around $150; multi-step workflows with AI or CRM logic run higher. You get the exact number on the discovery call.
Who owns the automation afterwards?
You do. Scenarios, workflows, code, prompts and the credentials created for the project are handed over with a walkthrough video and documentation. There is no lock-in and no hosting dependency on me.
Is the chat assistant on the site a real person?
No. "Max" is an AI assistant trained on these pages; it can explain services and cases and point you to the form. Alex personally answers everything that comes through the contact form or Telegram.
Describe your process
Which tools are involved, what is done by hand today, how often. I reply within 24 hours with a fit/no-fit answer and, if it fits, a fixed quote.