LLM products
An AI assistant in a mobile app on a custom agent harness: tool calling, streaming, memory, web search. Quotas, token accounting and prompt-injection protection.
- tools
- memory
- web search
- quotas
I build LLM products: agents, RAG and infrastructure for models. 10 years in backend development, 4 of them as an architect and team lead.
Saint Petersburg, working remotely
Ask my agent
This is how the agent harness I build into products works: a question, tool calls, an answer. The answers are prepared from my CV in advance, no model is called here.
I answer from Roman’s CV. Pick a question: first you will see which tools I call, then the answer itself.
Questions
What I do
I take LLMs to production: from a model on my own server to a feature people actually use.
Document recognition
An eval bench, a model cascade and a spending cap for 30k pages of PDF protocols.
53%%
accuracy on the reference set
Model infrastructure
An LLM gateway on LiteLLM with keys, budgets and fallbacks. Open-weight models on my own servers.
- LiteLLM
- vLLM
- Ollama
- Whisper
RAG and quality evaluation
Document search with hybrid ranking. I check answer quality against an eval set of questions.
- pgvector
- hybrid search
- eval set
Teams and architecture
10 years in backend with PHP/Symfony, Go and Python. Moved my team to agentic development.
%
faster task delivery
Projects
Underdogi
Team lead and AI engineer, 2026
LLM recognition of competition PDF protocols and fighter ratings.
10 s → 60 ms
rating page
Go-Base Platform
Agentic development, 2026
An ERP/WMS warehouse system on heavy legacy code, 58 business modules.
40%
faster tasks
GigWork
Architect, 2024–2025
An HRM platform for staff outsourcing: PHP/Symfony and Go microservices.
≈25%
faster feature delivery
Too Many Gifts
Architect and team lead, 2021–2023
A B2B store with over a million products, designed from scratch.
15+
suppliers in import