AI / LLM Engineer

Designing and implementing AI agents from scratch. Technical owner of the LLM layer across several client projects: RAG systems on Qdrant (embeddings, semantic search), model selection and routing, LLM API integration (Claude, Gemini, OpenAI) and local models (Qwen), answer-quality monitoring with Langfuse, delivery. Embedding model migration: answer quality on the knowledge base improved from 0.59 to 0.82.
Stack: Python, FastAPI, PostgreSQL, Qdrant, Langfuse, Docker, React

Designing and implementing LLM solutions: RAG systems on Qdrant, AI consultants and assistants with tool calling, Telegram bots with AI integration, cross-platform authentication, open-source data aggregation. 43% growth in bookings on one project.
Stack: Python, FastAPI, Qdrant, Langfuse, PostgreSQL, Docker, React, GitLab CI/CD, llama.cpp, Ollama.

Designing and implementing an LLM assistant for first-line support across a fleet of 10,000+ devices. The assistant autonomously resolves ~73% of L1 tickets (600–800 per month).
Prompt-layer design: system prompts, few-shot examples, response format. Structuring the knowledge base as model input. Testing answer quality and iterating on prompts. Photo-based recognition of working and faulty Raspberry Pi devices on the support line (computer vision). Fleet telemetry: requirements gathering, Grafana dashboards, Telegram alerting. Mentoring a mid-level engineer.
Stack: Python, LLM API, Linux, Grafana, Confluence, Jira.

Participated in rolling out ML models on the first line of support: gathering and formalizing requirements, testing models, working with the development team. SQL analysis of user sessions, Grafana alerting, REST API diagnostics, incident management. Regular monitoring of product health and performance.
Stack: Linux, SQL, Grafana, Jira, REST API.

Designing, implementing and maintaining production services on a VPS (Docker Compose, Caddy, CI/CD): AI agents and Telegram bots with Mini Apps, RAG on pgvector, cost-aware Claude model routing, self-hosted LLMs (llama.cpp, Ollama), monitoring with Langfuse.
Portfolio: heylark.dev.