Zoran Stojanovic

Senior Applied AI & Automation Engineer | Multi-Agent Orchestration & Deterministic Systems

Belgrade, Serbia • Remote Worldwide (Full US / European Time Zone Overlap)

🏆 Kaggle Top 1.2% Expert (#734 / 61,348) GitHub Developer Program Member

I architect and deploy production multi-agent workflows, deterministic LLM governance runtimes, and high-reliability machine learning systems. Creator of JudgeGuard, an open-source multi-tier verification engine turning probabilistic LLMs into zero-trust, reality-grounded software components. Deep engineering track record in WebContainer / in-browser dev environments (Bolt.new), autonomous coding agents, universal RAG engines (ChromaDB / Llama-3.3), and calibrated clinical risk models (MedGemma, Brier Score calibration).

Download Master CV (PDF) GitHub (60+ Repos) Kaggle Expert (#734) LinkedIn kiza101288@gmail.com

Production & Open-Source Engineering

JudgeGuard Core & Autonomous Multi-Agent Governance Engine Lead Architect

Designed and open-sourced an enterprise multi-tier governance runtime enforcing deterministic verification on autonomous AI agents before and after code execution.
Engineered Layer 00 (regex blocks on destructive actions), Layer 000 (Ollama Gemma 4 gatekeeper), Layer 3 (semantic drift detection under <0.40 threshold), and Layer 4 (physical reality verification inspecting disk state, SHA256 hashes, and non-empty byte outputs).
Eliminated mock hallucination loops and false positive completion claims, enforcing mandatory git checkpointing with SHA recording on every successful task.

Bolt.new & WebContainer Optimization Engine (git_bolt_jules_gate.py) 46k+ Tokens Saved / Run

Engineered an automated context-reduction and surgical diff generator specifically designed for browser-based IDEs (WebContainers / Bolt.new).
Developed token-cleaning protocols isolating unneeded cache directories (`node_modules`, `.cache`, `dist`), saving 46,000+ prompt tokens per session and preventing context overflow in browser runtimes.
Implemented surgical diff algorithms that replace multi-thousand-line full rewrites with targeted AST replacement blocks.

Autonomous Developer Agent v2.0 (Google Gemma 4 Competition) 43% Faster Execution

Engineered a single-agent autonomous coder running Gemma 4 31B, calibrated for competitive problem-solving across held-out, unseen software repositories.
Formulated a strict call-paced doctrine (Calls 1–8: semantic read, Calls 9–12: atomic edit, Calls 13–20: test verification, Call 20+: diff checkpoint), reducing execution latency by 43% while staying within a 12-hour air-gapped evaluation ceiling.

Clinical Risk Prediction Pipeline & Healthcare AI (MedGemma) Top 1.2% Ranked

Built an end-to-end clinical risk prediction pipeline, evaluating 15 model iterations to optimize the trade-off between Brier Score (probability calibration) and F1 score, preventing overconfident false positives in medical decision support.
Created MyHealth Decoder (MedGemma Patient Advocate) for the Google Med-Gemma Impact Challenge; achieved 0.9576 Apex score in the RSNA Knee Abnormality Detection medical imaging competition using triple-model fusion.

ResearchProAI (Usisivac-V6) — Universal RAG Ingestion Pipeline Production RAG

Built an autonomous knowledge ingestion system vacuuming codebases, API documentation, and engineering logs into a ChromaDB vector store.
Integrated Llama-3.3 for structured synthesis, accompanied by a live CLI dashboard monitoring knowledge saturation and chunk retrieval precision.

Diskort — Asynchronous Go Automation Suite 123+ GitHub Stars

Authored a high-throughput, concurrent automation and bot orchestration suite in Go, achieving 123+ GitHub stars with zero memory leaks under prolonged execution.