Backend systems
APIs, data models, async workers and the unglamorous plumbing that keeps a product reliable as traffic grows.
Senior Backend Engineer · Python & AI
Systems thinking and understanding the domain matter more than ever. 6 years on SaaS products: backends, data acquisition and AI-backed features. Let's build something stable and scalable together.
Products I've built and shipped
Backend-first, comfortable across the stack. I pick boring, proven tools and make them ship outcomes.
6 years of backend work on SaaS products, in fast-paced startups and in larger companies, across HR, AdTech, AI search and GEO, fraud detection and telecom. One thread runs through all of it: data acquisition. Crawlers, ETL pipelines and custom collection mechanisms for domains with no reliable structured source, from advertising fraud traffic to the answers generative AI gives about a brand.
AI coding tools are part of how I work every day, and I am deliberately growing into AI engineering. As code generation gets cheaper and faster, the bottleneck moves elsewhere: designing systems that hold together, understanding the domain well enough to know which problem is worth solving, making the right architectural trade-offs.
I hold a high bar for quality. Clean architecture, meaningful tests, readable code and performance that survives real load, because that is what keeps a system cheap to change a year from now.
Engagements built around shipping software outcomes - not billing hours.
APIs, data models, async workers and the unglamorous plumbing that keeps a product reliable as traffic grows.
Crawlers, ETL pipelines and custom collection mechanisms for domains with no reliable structured source, turned into a dataset you can actually query.
LLM-backed features and agentic workflows on LangChain and LangGraph, with retrieval over your own data so answers come from your catalog, not from model memory.
Versioned, documented, well-tested REST APIs that frontend teams and integrations actually enjoy using.
Query and endpoint optimisation, test coverage that catches regressions in CI, and refactors that make a legacy codebase safe to change again.
Architecture reviews, stack decisions and pragmatic guidance for teams scaling a Python codebase.
Project teardowns: the problem, the stack and the outcome.
AI platform for brand-visibility and traffic-quality analytics - with an AI Growth Coworker, GEO and AI-visibility audit modules, and traffic analysis that separates real visitors from bots and crawlers.
Medical-cannabis data platform for Poland: a strain catalog with terpene and THC/CBD profiles, near-real-time pharmacy availability and prices, prescribing clinics, and Dendi - an AI assistant (RAG) that checks local availability and suggests similar-profile alternatives.
Loose, educational projects - I build them to get hands-on with a new technology or to crack a specific problem.
Asking-price valuation for flats in Rzeszów - an ML model trained on sale listings from the city. From the address alone it enriches each record with a dozen extra location features generated from OpenStreetMap. Validation MAPE 6.7%, against 15.5% for a naive baseline.
A streaming-style movie recommender with two independent AI systems: personalised recommendations and hybrid semantic search, behind a Netflix-like UI that hides the algorithms.
Control your home by voice over Telegram, in plain Polish and with your own device names. A voice note is transcribed, an LLM turns it into a structured intent, and the system flips the relay locally over the LAN - with a hard guarantee it can't invent a device that doesn't exist.
Got a backend that has to scale, data with no clean source to pull it from, or an AI feature that has to be trustworthy? Drop me a line.
hi@krystianjarmul.dev