The lab · five practices

What I build, end to end.

Agentic systems in production, AI inside the dev loop, modeling under the hood, multi-cloud infrastructure, and the frontend that ships it. Five distinct practices.

Practice · I

Agentic AI systems

Production agents, RAG, and LLMs ingested into data pipelines.

Multi-step agent graphs with deterministic edges and generative middles. Hybrid retrieval over Qdrant, LLM rerank on the top-K, payload-based access control at query time so authorization can't be bypassed in the answer. Beyond chat: agents and LLMs embedded inside data pipelines — extraction, classification, enrichment, validation — wherever a rule engine bends and a model can fit. Every step traced, debuggable down to the exact tool argument.

LangGraphQdrantLangfuse v3tool callsRAGobservabilitydata-pipeline AI

Practice · II

AI-augmented engineering

Claude Code, Codex, MCP. AI inside the dev loop, not a sidebar.

Building skills, plugins, and MCP servers that turn an LLM IDE into a real engineering surface. Agent-driven workflows that plan, write, review, and ship code under supervision — without losing the human judgment that decides what's worth shipping. The output isn't faster typing, it's smaller teams shipping bigger products.

Claude CodeCodexMCP serversskillspluginshooksCursor

Practice · III

Modeling & Forecasting

Deep learning, classical ML, time-series. The math under the agents.

General data science depth — classical machine learning with scikit-learn and XGBoost on engineered features; time-series forecasting using both traditional methods and deep-learning approaches. LoRA fine-tuning of foundation models — Gemma, Llama 3.1 — when an adapter actually changes behaviour beyond what prompting can do. MSc in Data Science from Warsaw University of Technology, plus ongoing research at the Polish Academy of Sciences.

scikit-learnXGBoostPyTorchTensorFlowLoRAtime-series

Practice · IV

Cloud infrastructure

AWS · Azure · GCP. Production across all three.

Terraform from day one — no clickops. Azure: App Services, Logic Apps, Key Vault, AI Foundry. AWS: Bedrock, SageMaker, Lambda. GCP: Cloud Run, Compute Engine, Vertex AI, Pub/Sub. Model serving and inference platforms wired alongside the rest of the stack. CI/CD via GitHub Actions, with gates that actually fail on drift.

TerraformAzureAWSGCPVertex AIGitHub Actions

Practice · V

Full-stack delivery

FastAPI ↔ Next.js. Eval harness. Dashboards. Ready scaffolds.

FastAPI backend with typed contracts. Next.js frontend — customer-facing UI for end users, dashboards for analytics, admin views for the team running the system. One stack, one repo, one deploy. Eval harness wired to keep model drift visible week over week. Ready scaffolds that bootstrap an AI-powered product in days — the team gets a working app, not a notebook and a slide deck.

FastAPINext.jsPostgreSQLeval harnessdashboards

Got something to build?

A short email is fine. Tell me what you're trying to do and where you're stuck — I'll write back.

The full bio & channelsTalk to the lab→