Portrait of Shahin Gurbanov in sunglasses and a red and teal patterned hoodie, with a motion-echo effect

Shahin GurbanovProduct owner / AI-assisted engineering

I direct AI to build real products.
Here are the receipts.

Seven working systems, from a live ticketing marketplace to a real-time voice receptionist. Every case study shows the architecture, my decisions, the numbers, and what doesn't work yet.

case studies
7
live in production
5
with AI in the product
5
automated tests, all projects
3,200+

Selected work

Real systems, honest write-ups.

Open any card for the architecture, my role, the numbers, the problems I hit and what is still unfinished.

01 Start here
Live in productionClaude + local speech models

AI Hypnotherapist

A spoken AI consultation in a real clinician's recorded voice, then a personalised hypnotherapy audio session.

307 automated tests passing (4 skipped)

Case study Open live service
02
Live in productionGemini text and image

Genbooks

Describe a book, read a free AI-written preview, pay a fixed price, and get a finished EPUB, PDF and cover by email.

4 to 6 min total time for a free preview in real runs

Case study Open live site
03
Live in productionAI in the workflow, not the product

SocialField

A live UK events and ticketing marketplace with Stripe Connect payouts, waitlists, memberships and six unattended jobs.

233 unit tests passing in 2.5 seconds

Case study Open live site
04
Working system (local)OpenAI Realtime voice

AI Sales Manager

A voice AI receptionist that talks live, by speech, to guests on a live stream through an animated avatar.

36 ms to hand a ready connection to the next guest

Case study
05
In daily personal use (local)Gemini for applications

Personal AI Job Hunter

Searches 79 employer boards and a dozen platforms, ranks vacancies, writes tailored applications and fills the forms, but never presses Submit.

~1,960 automated tests

Case study
06
Live, running dailyGemini, quote-verified

Backlink Builder

A safety-first automation that finds guest-post opportunities, makes the model prove every decision with a quote, and runs unattended every day.

458 tests pass in the public showcase

Case study Open showcase
07
Live in productionAI in the workflow, not the product

Singles

A live video speed-dating platform that runs real, moderated events in the browser on Cloudflare's free tier.

122 automated tests in 17 files

Case study Open live app

At a glance

What is actually inside each project

A recruiter's shortcut: which projects contain an LLM, retrieval, voice, payments or unattended jobs.

Project LLM in the productRetrievalSpeech and voiceTool callingPaymentsBrowser or UI automationUnattended jobsPublic and live
AI Hypnotherapist YesClaude Opus 5.5 YesBM25, no embeddings YesLocal TTS and STT No YesStripe Checkout No YesBackground loops Yes
Genbooks YesGemini 3.1 Pro No No No YesStripe Checkout No YesWatchdogs and cron Yes
SocialField No No No No YesStripe Connect No Yes6 scheduled jobs Yes
AI Sales Manager YesOpenAI Realtime YesEmbeddings and vector index YesSpeech to speech Yes3 tools No YesWindows UI Automation YesAuto-start listener No
Personal AI Job Hunter YesGemini No No No No YesPlaywright No No
Backlink Builder YesGemini (free tier) No No No No YesPlaywright YesDaily GitHub Actions Yes
Singles No No No No No No YesDurable Object alarms Yes

Taken from each project's own report. A dot means the feature exists in the code; a dot with a note names it. Blank means it does not exist. SocialField and Singles contain no AI: AI is how they were built.

How I work

Five habits that show up in every project

These are not slogans. Each one links to the place in a case study where it actually happened.

01

Measure first, then decide

Cost, latency, capacity and funnel losses get counted before anything is fixed or sized.

  • AI Hypnotherapist load-tested with simulated clients, then set the waiting room at 3 concurrent consultations.
  • AI Sales Manager found the AI's voice was 73% of cost, then cut a 3-minute guest from about $0.065 to about $0.04 (estimated from three real chats).
  • Backlink Builder counted where pages were lost (about 410 opened, 3 passed in one week) before changing anything.
02

Guardrails live in code, not only in prompts

A prompt can be ignored. A validator can't.

  • Genbooks a code gate rejects AI-sounding text and triggers a repair pass; truncated output is never saved as complete.
  • AI Hypnotherapist every client message passes a keyword net and an independent AI classifier, and the system fails safe if the classifier is down.
  • Backlink Builder every model 'yes' needs an exact quote, and plain code checks the quote is really on the page.
03

A human stays where it matters

Automation does the repetitive work. People keep the decisions, the CAPTCHAs and the final button.

  • Personal AI Job Hunter never presses Submit. It fills the form and hands the bot check to the person.
  • AI Hypnotherapist an optional clinician review before a script is released, and a human practitioner as the alternative.
  • Backlink Builder never touches a CAPTCHA. A Chrome extension fills the rest while the owner ticks the box.
04

Every bug becomes a test

Defects found in real use are reproduced, fixed and pinned down so they stay fixed.

  • Backlink Builder each named regression test fails on the pre-fix code (checked at the commit before each fix).
  • Singles old bugs were put back one at a time to prove the tests fail.
  • Personal AI Job Hunter duplicate detection was checked by replaying the real history of 923 sent and skipped jobs.
05

Say what it isn't

Every case study lists what is not built and what can't be verified, so nothing is oversold.

  • SocialField states plainly that there is no LLM in the product. The AI is in how it was built.
  • Backlink Builder no backlink is confirmed live yet, so none is claimed.
  • AI Sales Manager lists what is not built, such as mid-guest spend-limit enforcement and a self-echo detector.

About

I own the product. AI helps build it.

I'm a hypnotherapist and events organiser in London who ships software by directing AI coding assistants. I set the requirements, make the architecture and cost decisions, test with real people and run the result in production. I don't claim to have hand-written the code, and each case study says exactly where that line sits.

More about how I work

Contact

Got a role or a project in mind?

If something here looks relevant to what you are building or hiring for, get in touch. I can walk you through any architecture or show it running.

shahingurbanov@yahoo.com+44 7404 073261