Vedran Balagović

EU-sovereign compliance engine for the EU AI Act. Self-host it, point it at your LLM, and it becomes the regulator's proof layer: a tamper-evident HMAC hash-chained audit log, one-click Annex IV dossier export, an Article 73 incident workflow, and PII / prompt-injection / bias checks. Multi-tenant, multi-LLM.

Local speech-to-text with speaker diarization — no Python, no cloud, a single binary. Wraps a pure-C Qwen3-ASR engine (PyTorch-matching quality, hand-tuned SIMD, CUDA) with Pyannote 3.0 diarization via sherpa-onnx, plus optional local-LLM speaker identification.

Used-car valuations computed from 266k+ live listings across 11 European markets, refreshed 4× daily — with photo-based AI condition checks and fair-price ranges. Powered by the CarsDataset API (370+ brands, CSV/SQL export). A study in scraping, normalization and data engineering at scale.

A 60-second daily check-in app for parents of kids 4–14; an AI companion surfaces patterns in mood, sleep and behavior and suggests evidence-based techniques. Flutter, iOS & Android, 10 languages — built and shipped to both stores solo.

An AI character-animation engine — turns inputs into character animations through an AI-driven pipeline, for creators and studios who want to skip manual animation work.

On-device Qwen3-ASR transcription with speaker diarization (gospeech) — pure-C inference engine, SIMD/CUDA kernels, streaming with encoder cache, and local-LLM speaker identification. PyTorch-matching quality, fully offline.

Retrieval-augmented generation systems end to end — document chunking, embeddings, vector search, re-ranking and grounded, cited answers. The plumbing that makes LLM features trustworthy in production.

MCP servers (Neural Draft), an LLM compliance engine (Lex Custis), and Promptual — a GPT-4o prompt-improver browser extension. The tooling that lets AI agents act safely and usefully.