Your AI worked on a demo. Production is eating your team alive.
You built something that looked great in a demo. Now two engineers are babysitting prompts, tokens are burning budget, and nobody can debug it when it breaks.
This isn't an AI problem. It's a systems engineering problem. I fix those.
Sound Familiar?
Two engineers fixing prompts instead of shipping features
AI was supposed to save time. Instead it's consuming your best people. They're not building product — they're nursing a system that should run itself.
"Works great on demo, hallucinates in production"
The gap between demo and production isn't a prompt problem. It's a missing feedback loop, no evaluation pipeline, and architecture that can't handle real-world input variance.
Token bill 10x what you budgeted
You budgeted $500/month, burned through $5K. The model is doing work that caching, routing, or a simpler architecture would handle at a fraction of the cost.
AI broke. Nobody knows where to look.
Logs exist but tell you nothing. No structured traces, no evaluation metrics, no way to distinguish between a model issue and a data issue. Black box instead of a system.
"ChatGPT can do it" ≠ "we can productionize it"
ChatGPT showed great results in a browser. Your API in production outputs something completely different. The gap isn't the model — it's everything around the model.
If you recognize even one of these — we should talk. I solve exactly these problems, systematically.
How I Actually Help
I don't sell hours. I sell outcomes. Two formats depending on what you need:
AI Systems Architecture
I design the system that makes your AI actually work in production. Not a strategy deck — a concrete architecture with implementation path.
What you get:
- Architecture with clear component boundaries and failure modes
- Observability and evaluation strategy built in from day one
- Integration roadmap that respects your existing systems
- Risk assessment with specific mitigation for each risk
Good fit if:
- Your demo works but production is unstable or expensive
- You're building an AI product from scratch and want to skip the expensive mistakes
- You need someone to tell you honestly what's wrong with your current architecture
- Legacy system + AI integration feels like defusing a bomb
Fractional Technical Advisor
Ongoing senior engineering judgement without a full-time hire. I embed into your decision-making process and catch problems before they become expensive.
What you get:
- Architecture decisions reviewed before they're committed to
- Critical technical calls with someone who's seen the failure modes
- Code and system review with actionable feedback, not academic theory
- Hiring support: what to look for, what questions to ask, which roles you actually need
Good fit if:
- Startup without a CTO or senior systems person
- Growing team making architectural decisions that will be expensive to reverse
- Building AI product and need an experienced advisor who won't bullshit you
What Clients Say
RAG in our chatbot worked just terribly. We tried to optimize it in every possible way — all to no avail. At first I was hostile towards Ivan — in the first session he asked a lot of questions, and what we didn't have time to discuss, he asked me to fill in as answers in a checklist. While I was doing this myself, I started to guess that the problem wasn't with RAG, but that it simply wasn't needed for our task. In further work with Ivan, we rebuilt the system into a simpler one. As a result, it started working more accurately (according to tests) and became cheaper to maintain. Most importantly — we began to understand how it works, where to look if something breaks, and how to expand the system.
Our team was building an agent to process customer requests from mailboxes. At first everything worked fine, but then we found ourselves in a state of constant support of the bot itself and re-checking of formed requests. We were about to abandon it completely, but we would have had to hire more customer relationship managers. We turned to Ivan for consultation. Ivan looked at our system, talked with the team, and a few days later brought, as he modestly said — "an approximate PoC of how it should work." This PoC as is has been working for us for the third month now with practically no complaints.
We needed to develop an AI module for our old backend that's about 6 years old. We tried ourselves, but it worked with too large delays on requests, and it wasn't clear at all how it would live on. Ivan immediately said that it couldn't be done this way, and started asking uncomfortable questions. Honestly, I thought he was just stalling for time. But when he showed the diagram of how the integration should be — everything fell into place. We rewrote the AI service in a month, and now the system works stably. Tokens cost us about $400/month, although we had budgeted three thousand for this load.
Selected Projects
Production Agentic AI System @ Monite
Multi-agent AI system for financial platform with multi-stage pipeline, structured logging, and Schema Guided Reasoning
Technologies: Python, FastAPI, Pydantic, OpenAI, PostgreSQL, PgVector, semantic-router, Kubernetes, SGR
Intelligent Document Processing Pipeline
Async OCR service with CV preprocessing and LLM extraction: manual 2-3 day processing replaced with sub-minute automation
Technologies: Python, FastAPI, PydanticAI, PostgreSQL, AWS S3, OpenCV, Docker
AI Insurance Document Processing Pipeline
End-to-end automated insurance document processing: claim processing time from 3-5 days to under 10 minutes
Technologies: Python, FastAPI, Pydantic, PostgreSQL, Docker, LLM, SGR
Enterprise AI Onboarding Platform
RAG-based intelligent onboarding system: average onboarding time from 3 months to 1 month
Technologies: Python, FastAPI, langchain, LLM, ChromaDB, fastembed, PostgreSQL
Quint Code — Decision Engineering for AI Agents
Open-source MCP tool that adds structured decision-making to AI coding agents. Frame problems, compare fairly, record decisions as contracts that know when they're stale.
Technologies: Go, SQLite, FPF, MCP Protocol
Is This You?
You're my ideal client if:
✓ Tech Lead / Engineering Manager
Your team built an AI feature that now requires more support than the rest of the product combined
✓ Founder with a working demo
Investors are excited but production keeps breaking and you can't figure out why
✓ CTO / Technical Director
You need someone to look at your AI architecture with fresh eyes and tell you what's actually wrong
When to reach out:
- ✓ AI system is live but unstable — and your team can't stabilize it
- ✓ Token costs are out of control and you don't know where the waste is
- ✓ You can't debug AI failures — it's a black box
- ✓ Legacy system + AI integration — afraid one wrong move breaks everything
- ✓ Team is drowning in AI support instead of shipping features
- ✓ You need an independent architecture review before committing to a direction
- ✓ Building from scratch and want to get the architecture right the first time
When NOT to reach out:
- ✗ You want a "magic prompt" that solves everything — I do engineering, not incantations
- ✗ You need it done yesterday with no regard for quality — speed without architecture creates the exact problems I fix
- ✗ You need validation, not diagnosis — I'll give you an honest assessment, even if it's uncomfortable
Let's Talk
Short form, 2 minutes. Describe your situation. I'll respond with whether I can help and what format makes sense.
Or reach out directly: zakutnii.ivan@gmail.com
How It Works
- Discovery Call (30 min)
You describe the situation. I ask uncomfortable questions. We see if there's a fit. - Scoped Proposal
I send a specific plan: what I'll do, what you'll get, what it costs. No vague retainers. - Work Begins
Architecture, implementation, advisory — whatever the situation requires. Typical first deliverable within 1-2 weeks.
What Happens After the Call?
If there's a fit, I send a concrete proposal within 2-3 days.
If I can't help, I'll tell you honestly and point you to someone who can.
P.S. If you're not sure whether your problem is something I can help with — reach out anyway. The most valuable consultations start with "I'm not even sure what's wrong."