Kaan Kuzu

Eindhoven University of Technology (TU/e)Jul 2029 (expected)

B.Sc. Computer Science and Engineering | Eindhoven, Netherlands

Availability: part-time during the academic year (8–12 h/week) from Nov 2026; full-time from Jul 2027.

Lycée Saint-Benoît2026

High school diploma, grade 87/100 | Istanbul, Turkey (French-Turkish bilingual program)

Honors and Awards: AP Calculus BC score 5/5; IELTS 7.0; DALF C1

Additional Coursework: Sabancı University Summer School (2024): Artificial Intelligence and Machine Learning, Introduction to Web Programming

Technical Skills

Projects

AudienceLabs | Multimodal audio/video analysis pipeline (Python, FastAPI, PyTorch, Modal)Apr 2026 - Present

Link: audiencelabs.net

Built a pipeline that ingests recordings of up to 4 hours, extracts audio with ffmpeg, transcribes it with word-level timestamps (Faster-Whisper) and aligns speaker diarization (pyannote 3.1) and audio-energy/emotion signals (librosa, wav2vec2) into a per-second feature timeline.
Designed a two-pass LLM analysis (segmentation, then event detection) followed by deterministic validation rules (minimum duration, truncation and mid-sentence checks, sentence-boundary snapping) and IoU-based deduplication, turning noisy model output into consistent results.
Implemented a pluggable scoring interface with a heuristic baseline and an XGBoost regressor, plus a weekly retraining job with seeded 80/20 split and a versioned model registry with automatic fallback; released the scoring stage as an open-source package with tests and CI (github.com/kaankuzu1/clip-virality-scorer).
Runs as fail-soft, idempotent GPU jobs on Modal with per-stage status in PostgreSQL and a Next.js dashboard; documented with ADRs. Sole developer, 120+ commits; open for beta requests.

Multi-agent content automation for a law practice | TypeScript, Claude API, Trigger.dev, SupabaseSep 2026 - Present

Building a daily workflow, now in final testing, in which six specialized LLM agents (research with legal-database tools, writing, design, captioning) prepare an informational post, with every step gated by the lawyer's approval over WhatsApp.
Added output validation per step, event and cost logging in PostgreSQL, and a watchdog that checks every 5 minutes for stalled or timed-out runs and resumes them from the last completed step.

Education