Kirill Fisyukovich
About
I am a Python Backend Developer from Brest, Belarus. I like building systems that are simple enough to operate, but strong enough to survive real production traffic: APIs, admin tools, integrations, data pipelines, background workers, and AI features with clear boundaries.
I enjoy the engineering space where backend code meets the physical world: networking, hardware-adjacent products, computer vision, video processing, and operational tooling. I have worked on restaurant platforms, media workflows, ClickHouse analytics, Telegram bots, parsing systems, payment features, and computer vision for buffet hardware.
What I Care About
Reliable backends. I prefer explicit data models, readable migrations, predictable queues, useful admin panels, and logs that explain what happened.
Python as a production tool. I like FastAPI, Django, PostgreSQL, ClickHouse, Celery, RabbitMQ, Redis, SQLAlchemy, Docker, and the quiet discipline of making these pieces easy to debug.
AI without fog. I am interested in AI systems that are treated like real software: typed contracts, evaluation, state machines, retrieval traces, latency budgets, and human-readable failure modes.
Practical speed. Fast systems are not only about benchmarks. They are about fewer moving parts, fewer network hops, good database choices, and code that does not surprise the next person who reads it.
Interests
Outside of day-to-day backend work, I like exploring networking, Linux tooling, computer vision, backend architecture, product operations, and the strange edge where software starts controlling real devices. I also enjoy writing technical notes because it forces vague ideas to become precise.
Currently
I am focused on Python backend engineering, AI infrastructure patterns, admin tools that make products easier to operate, and systems where performance and maintainability matter at the same time.
Latest Posts
-
Exactly-Once Is a Product Lie: Idempotent AI Workflows in Postgres
2026-05-24
A serious pattern for AI workflows: use Postgres to make model calls retryable, auditable, and recoverable without pretending distributed systems are exactly-once.
-
Capability Security for LLM Tool Use
2026-05-23
Tool-use becomes safer when the model receives limited capabilities instead of global tools. Permissions are delegated, attenuated, validated, and logged.
-
RAG as a Query Planner, Not a Vector Search Feature
2026-05-22
Interesting RAG is closer to a query planner than a semantic search box. SQL, graph edges, vector similarity, reranking, and eval traces each do different work.