Local models and agents
I design environments for working with code: model runtimes, tools, editing constraints and result verification.
AI in products.From scope to release.
I design AI features and connect models to data and real workflows. I own requirements, UX, integrations and release quality.
Explore the work01 / Skills
What I build with AI
Knowledge-based assistants
I ground answers in project documentation and configure retrieval, context, language, persona and assistant behavior.
Support automation
I connect LLMs, external APIs and databases for multi-step support, profile verification, tickets and feedback.
02 / Projects
Where it is implemented
LOCAL / CODING AGENT
- Find the relevant code
- Apply a bounded edit
- Verify the result
Qwen · PHP / Python · Docker
INTERNAL ENGINEERING PROJECT
Local AI coding agent
An environment for PHP and Python repositories. I designed the architecture, tool integration, editing rules, verification and model routing.
- internal scenarios with PASS
- 7
- increase in tested context
- 4.91×
Internal tests · 13–14 September 2026
Case study, demo and measurementsSUPPORT / AI ASSISTANT
- Ban / appeal
- Technical issue
- Player report
- Purchase
FAQ → support request → ticket
RAG / FAQ / SUPPORT WORKFLOWS
AI support assistant
Knowledge-based answers, multi-step support, Steam profile verification and ticket creation through the existing Discord support module.
- Python modules in the AI subsystem
- 11
- response languages: RU / UA / EN
- 3
6.1K+ Python lines · AI subsystem snapshot
Commercial AI RAG · X-BOT caseInside the AI subsystemRetrieval, data, behavior and operations
Multi-step support flows
Four flows: ban appeals, technical issues, player reports and purchases. Intent detection, field extraction, multi-turn state and ticket creation.
Languages and behavior
RU/UA/EN detection and corrective regeneration if the output language is wrong. System instructions, tone, boundaries and context rules define assistant behavior.
FAQ and feedback
SQLite stores server configuration and static FAQ; MySQL stores dynamic FAQ and feedback. Schema/index creation, usage counters, reaction-based answer promotion and manual-review flags.
Retrieval and source grounding
Retrieval thresholds, top-K, context limits, documentation chunking and source tags. A groundedness score checks source IDs and lexical overlap between the answer and retrieved context.
Steam profile verification
Profile lookup, SteamID64 and vanity-URL resolution, followed by user confirmation. Structured profile data enters the support flow before ticket creation.
Configuration and operations
Per-server channel, trainer role, system prompt and MySQL settings. FAQ CRUD/reset, database connection tests, connection pooling, cache invalidation and runtime status.
03 / Process
How I take work to release
AI contributes to implementation. I retain ownership of product logic, UX decisions and acceptance of the result.
Scope and UX
Define priorities, user workflows, constraints, edge cases and acceptance criteria.
Implementation
Prepare context and tasks for AI. Direct edits, debugging and iterations against refined requirements.
Verification
Manually test features, interfaces and integrations. Check regressions, language and the grounding of AI answers.
Release
Validate the environment, deployment and rollback readiness. Accept the result against completion criteria.
Ready means verified
Release acceptance covers functionality, edge cases and regressions. AI features also require checks of language, sources, behavior constraints and the support workflow.
PRODUCT EXPERIENCE AND TECHNICAL DEPTH