I build web and AI products in TypeScript and Python, and I use AI coding agents every day. Writing code got cheap, so my effort goes into checking it: I read every diff, keep tests and automated checks around the agent, and choose models with my own evals. That's how I run pripremi.rs in production, and how I built Fulcrum, 2nd place at Adapty's First Commit hackathon. I co-founded GrowthQ Lab, a small studio where I build and run my own products, so I think about what a feature costs and who it's for, not only how to build it. I have a BSc in Computer Science and I'm looking for full-stack or AI engineering roles in Serbia or remote.
selected work
pripremi.rs
live SaaS
Lesson plans for Serbian primary school teachers, generated in about 2 minutes. I built and run all of it: generation across OpenAI and Gemini with a quality gate and failover, the API, local card payments through OTP Bank (in final testing), and distribution through teacher communities. My own eval harness picks the models on cost per valid document, and 2,700+ tests run on every change, plus a live smoke test against production.
A guard layer that lets an AI agent work with Adapty A/B tests without being able to call a fake winner. The agent's only write path is a TypeScript MCP server that checks every action against a policy, a sequential stopping boundary and a human approval, and logs it to a ledger.
Turns Claude Code and Codex session transcripts into a local, plain-markdown knowledge base and serves it back to the agent mid-session over MCP, so past decisions get consulted instead of rediscovered. Listed on awesome-claude-code, with 1,400+ tests, 70+ npm releases and an Obsidian plugin in the community directory with 180+ downloads.
A public benchmark for how well language models handle Serbian: grammar, pravopis, the aorist, jat, č/ć and Cyrillic/Latin consistency, the things translated English tests can't check. The official set has 351 items checked against Pravopis 2010, 30 models are on the leaderboard, and I opened a PR to add it to EleutherAI's lm-evaluation-harness (not merged yet). After public feedback that it had become too easy, v2 added 76 verified whole-passage corrections.
Python · lm-evaluation-harness · pytest · OpenRouter · Hugging Face
GrowthQ ↗shippedAn AI content product from GrowthQ Lab, the studio I co-founded in 2025: branded social media content generated from a per-brand profile. Selected for NTP Niš Launcher (7 of 108 teams), grant-funded, and pitched to investors at Web Summit Lisbon (2025) and GITEX Europe Berlin (2026). A US-market version is in development. React · TypeScript · Express · MongoDB · OpenAI · Kie.ai
sterijashippedA bilingual site with Maru, a Claude-powered chat assistant that reserves its worst-case cost before answering, keeps a per-IP daily budget and rate limit, and is checked by an eval suite. Astro · Cloudflare Workers · Claude API
Fokus ↗shippedAn always-on-top desktop focus timer with a floating widget and global-shortcut capture, released in 15 versions for macOS and Windows. Tauri · Rust · React · SQLite
anglerexperimentA photo-culling tool I'm building for The Next Angle, a photo and video studio I invest in. An offline Python CLI scores shots on faces, sharpness, exposure and bursts and writes Lightroom ratings as XMP sidecars. Python · OpenCV · ONNX Runtime
klinexperimentTwo Claude Code plugins: a task memory loop that keeps what an agent learned from slipping away at /clear, and model routing between Sonnet, Opus and Fable. I use it on two products; no benchmark yet. Claude Code plugins · Shell
serbretrievalexperimentA Serbian search benchmark that measures how results change when the same query is written in Cyrillic, Latin, or Latin without č, ć, š, ž and đ. The tooling and a first corpus of 7,009 passages are done; queries and grading come next. Python
TRAINexperimentAn AI fitness and nutrition app: training and meal plans, a weekly check-in that adjusts calories from weight and logged meals, and web push reminders. Built and tested on test data, not launched. Next.js · Supabase · Claude
digital twinresearchIn my free time I'm researching a digital twin: AI that knows how I work and asks me the questions I'd rather avoid, instead of giving advice. One part is a private chat app, another a small experiment where an agent questions me by written rules. Next.js · SQLite · Claude API