Hi, I'm

Uroš Pantelić

Software Engineer — Backend & ML Infrastructure

Electrical Engineering graduate (University of Belgrade, Aug 2026) with industry experience spanning backend systems, ML infrastructure, and large-scale distributed platforms. Interned at Microsoft and Tenstorrent working on LLM inference and training systems; built production backend services and full-stack applications independently. Passionate about building reliable, scalable software and solving hard engineering problems end to end.

GitHub Download CV
  • 📍 Belgrade, Serbia
  • 🎓 University of Belgrade — B.Sc. Electrical Engineering, GPA 9.0/10.0
  • 🗣️ Serbian (native) · English (C1) · Turkish (basic)
UP

Experience

Nextesy

May 2026 — Aug 2026

Software Engineer Intern, Backend

  • Designed and implemented a role-based permission system for a multi-tenant backend, supporting fine-grained access control across user roles, resources, and organizational scopes.
  • Optimized permission resolution through indexed DB queries and query batching to minimize authorization overhead on high-traffic endpoints.
  • Built an integration and load testing suite that verifies access-rule correctness and benchmarks endpoint throughput under concurrent load.

Tenstorrent

Feb 2025 — Nov 2025

Machine Learning Intern & Open Source Contributor (TT-XLA)

  • Enabled parameter-efficient fine-tuning (LoRA/DoRA) for LLaMA 3.2 1B by extending training pipelines for scalable multi-device adaptation.
  • Built reproducible training/evaluation pipelines; benchmarked latency, throughput, tokens/sec, and memory utilization across device configurations.
  • Implemented tensor-parallel inference for LLaMA 3.1 8B in JAX/Flax, compiled through StableHLO into Tenstorrent's MLIR-based compiler stack (TT-XLA); validated numerical parity against PyTorch/HuggingFace baselines and contributed upstream.

Microsoft

Aug 2024 — Nov 2024

Machine Learning Engineer Intern

  • Researched selective state-space models (Mamba) and built internal prototypes in PyTorch; constructed a benchmarking harness against transformer baselines (LLaMA/GPT-style).
  • Benchmarked latency, throughput, memory usage, and quality metrics; delivered a technical report that informed model selection and deployment trade-offs.

Competitions & Hackathons

🥉
Qualcomm LPCVC 2025, Track 1

3rd Place — image classification under real-world distribution shifts, deployed on Qualcomm AI Hub (Snapdragon 8 Elite QRD). Presented at CVPR 2025 (Nashville).

🥈
DigiHack Hackathon

2nd Place — predictive maintenance for industrial equipment using sensor data and anomaly detection.

🥈
ASEF Hackathon

2nd Place — air pollution forecasting system for Belgrade using weather-based features and regression models.

Projects

Cloud-Native Distributed Inference Platform

PythonFastAPIWebSocketsDockerRedis
  • Distributed job orchestration system for CNN inference, modeled on cloud-native scheduler architecture (Kubernetes-style job queues with pluggable executors).
  • Worker pool management with dynamic scaling, DAG-based dependency resolution, task prioritization, and automatic fault recovery with configurable retry logic.
  • Real-time observability dashboard (WebSockets) exposing per-node load, task throughput, queue depth, and retry rates; near-linear scaling under varying concurrency.

AI Agent Orchestrator 2026

PythonClaude CodetmuxGit
  • CLI and daemon that runs multiple LLM coding agents in parallel, each in its own git worktree and tmux session, pulling tasks from a shared queue with per-task locking.
  • Full task lifecycle (dispatch, park, resume, cancel): parking commits work-in-progress and releases the lock so a task can resume on any idle worker without losing state.
  • Deterministic recovery from outages and self-healing of leaked locks, plus a web UI board and per-worker logs for monitoring running agents.

Study Feed 2026

FlutterDartRiverpodDriftClaude API
  • Mobile app that turns course materials (PDFs, text, images, YouTube links) into summaries, quizzes, and short educational character dialogs, studied through a vertical TikTok-style feed.
  • Content pipeline built around the Claude API and ElevenLabs text-to-speech, with word-synced captions timed from the TTS output and cached background video loops.
  • Local-first with a Drift (SQLite) database and a spaced-repetition review ladder with streaks; optional Google sign-in backs up progress to Firebase.

Find Me Around Belgrade

An interactive map of where I actually spend my time. Click a pin, or click a card.