AI systems in production
Lviv, Ukraine
Where the demo ends.
I test AI systems before your customers do.
Not ready for a call? Checklist: 11 quiet failure modes of AI agents (PDF)
- 7 → 9 of 10correct bot answers after fixing retrieval, prompt untouched
- 3 days → 1for code review after one repository setting
- Twilio → our ownvideo and voice calls on custom WebRTC I built as CTO. Two years in production
“He never throws ‘AI magic’ at a problem. He understands data, constraints, operations, and risks. His solutions are stable, explainable, and maintainable. When you need a system that works today and still works six months later, he’s the person you trust.”
What I’ve built
- CTO of a live consultation platform. Chat, voice and video on custom WebRTC instead of Twilio: 1,000+ advisors, 100+ concurrent sessions, a team of 5 engineers, two years in production.
- A healthcare RAG bot. 10,000 users, 1,000+ concurrent conversations at peak. Multi-tenant, with role-based access.
- An AI sales assistant for B2B. 20 client companies on one platform, responses in under a second, LLM costs down 45%. Led a team of 4 engineers.
- A tool for marketplace sellers. 10,000 sellers, 1,000 product updates a minute: ranking, reviews, conversion.
- An HR system for a clinic. 200 staff, integrated with the clinic’s medical records system. Tech lead of a team of 5.
Python · FastAPI · PostgreSQL · AWS · LangChain and LangGraph · vector databases · React
When people call me
- Everything works in the demo, but in production the bot gives customers wrong answers, and nobody knows how often.
- The model bill keeps growing, and nobody can see what exactly you’re paying for.
- The tests are green, but nobody can say what exactly they prove.
- The team writes code with AI faster than it can review it.
- You have to decide whether to build or buy, and a wrong call will cost months.
- You learn about failures from your customers, not from monitoring.
Sound familiar? Tell me what isn’t working.
Book a call · 30 minWhat it seemed, what it turned out
Seemed:
the bot needed a better prompt.
Turned out: it was retrieving the wrong pieces of the documents. We changed how documents were split and the retrieval order, and didn’t touch the prompt. On the test questions the bot started answering correctly 9 times out of 10 instead of 7.
health startup, RAG
Seemed:
the team was slow at writing code.
Turned out: code waited days for review. One repository setting helped, not a new process. Review went from three days to one.
my current team, under NDA
Seemed:
delivery sales had been falling for five months.
Turned out: at one outlet the till booked 12 of 32 marketplace orders as payments on the chain’s own website. The channel was falling less than the system showed. The marketplace portal also showed both outlets offline for 12–14% of scheduled hours, about double the previous half-year.
café chain, point-of-sale data analysis, pilot
“He introduced a judge node to our agent architecture, which brought a meaningful improvement to output quality.”
What I do and what it costs
from $2,000 · one system
Express audit of one system
One bot, RAG system or agent. 30–50 checks on your real queries, retrieval and answers checked separately. A report: what was checked, what wasn’t, and what to fix first. A 45-minute walkthrough. Takes 3 working days. Sample report (PDF, fictional data).
from $5,000 · up to three components
Full product audit
Up to three components. 100+ checks your team keeps. Cost and response time, exposure to instructions hidden in your data, access rights, monitoring gaps. A 30/60/90-day plan and a walkthrough with your team. Takes about 2 weeks.
from $3,000 · your team’s workflow
AI in your team’s workflow
How your team writes code with AI: rules for agents, reviewing AI-written code, checks in CI. Interviews with 2–3 developers and a list of changes, so AI speeds the team up instead of adding debt. Takes about 1.5 weeks.
priced after scoping · for your task
Building an agent, RAG system or automation
For a specific business task. First we agree how we’ll measure success: on which examples, with what accuracy, cost and response time. Then I build and test it before launch. Takes from a few weeks.
Right now I’m running a pilot with a coffee-shop chain: an agent works with their point-of-sale data and helps find where money gets lost. First findings are in, one case below.
I give the exact price after a 30-minute call, once it’s clear how many systems are involved and what access I’ll have.
If we start a full audit or a build within 60 days of an audit, its fee is credited. So if you’re not sure where to start, start with the express audit.
A re-check of your fixes 2–4 weeks later is included in every audit.
The first three audits are 30% off in exchange for a testimonial and an anonymised case study.
How we work
A 30-minute call
Free. You tell me what isn’t working or what you want to build. I tell you honestly whether I can help.
A proposal with no surprises
Price and timeline are fixed before we start. If the scope grows, we talk first, then re-price.
The work
An audit takes 3 working days to 2 weeks. A build follows the agreed plan.
Walkthrough and next step
A report and a walkthrough with your team. Then your team fixes things on its own or with me, in which case the audit fee is credited.
Who it’s for
A good fit if
- Your AI feature is in production or about to be, and you need to know whether you can trust it.
- Your team writes code with AI and doesn’t want to pile up debt.
- You need an agent or RAG system for a specific task, with results you can measure.
Not a fit if
- You need a slide deck about AI strategy with no system to verify.
- You need a quick demo that isn’t meant to go to production.
- You need a promised result before anything has been checked.
Questions
What exactly is in the report?
A table of checks with the result of each. Real queries where the system gets it wrong, and the cause behind each type of error. What wasn’t checked and why. What to fix first. The report comes as a PDF, with the checks your team can rerun after every change and the plan as a CSV for your task tracker. See the format: sample report (PDF; the company and data are fictional).
Do you work under NDA?
Yes. Most of my projects are under NDA, which is why there are no client names here.
Who does the work?
I do, myself, with no subcontractors.
What access do you need? Do you see my customers’ data?
The minimum: access to the bot or API in a test environment, 30–50 typical queries (anonymised is fine) and logs, if you have them. I don’t need access to your customers’ personal data. If the data is sensitive, we sign the NDA before we start, and I can work inside your infrastructure.
Remote? Which language?
Yes, remotely, I’m based in Lviv, Ukraine. English or Ukrainian.
What if the audit shows everything is fine?
You get a report of exactly what was checked, plus checks your team keeps and can run after every change.
What if the problem isn’t AI?
I’ll say so. In one of the cases above, a repository setting helped, not a new model.
Mentoring
An engineer growing into AI? A review of your AI project, or Career Inbound, an 8-week course I run with recruiter Oksana Holubets (in Ukrainian).
Tell me what isn’t working.
Book a call · 30 minI write about what breaks: the Telegram channel “Why it breaks” (in Ukrainian) and LinkedIn.