Veterinarian turned speech-pipeline engineer. I build the tooling that turns raw audio into broadcast-quality text — ASR alignment, forced alignment, subtitle automation. Based in Egypt, working remotely.
I came at this from the other end: 50,000+ professional captioning and subtitling projects delivered across 7+ languages before I wrote a line of pipeline code. That's why my tools optimize for the things that actually break in production — line balancing, reading speed, where a caption should not split — rather than for a benchmark number.
Word-accurate subtitle timing from audio + transcript, or from audio alone.
Multiple speech engines cascaded so each one only handles what the previous couldn't solve — Vosk on CPU covers the bulk, WhisperX + wav2vec2 adds GPU precision, Montreal Forced Aligner runs surgically on the leftover gaps, and interpolation closes the rest. Output goes through a 3-phase iterative segmenter that enforces real captioning constraints (32 CPL, 3s max duration, clause-aware line balancing across 7 languages).
With no transcript, the engines vote: a time-anchored, confidence-weighted consensus in the spirit of ROVER, with WhisperX as the timing backbone and Parakeet TDT / Canary-Qwen as high-accuracy challengers. Every consensus word keeps a fused agreement score, so low-confidence regions can be flagged for human review — in professional captioning, knowing where the transcript is uncertain is worth as much as the transcript.
Python · WhisperX · wav2vec2 · MFA · Vosk · NeMo · FFmpeg
- Speech & alignment pipelines — ASR, forced alignment, timestamp recovery, multi-engine ensembles, subtitle/caption generation at scale.
- Local LLM deployment — quantization and inference optimization on consumer GPUs, plus multi-provider cloud API orchestration with failover.
- Media automation — multilingual video pipelines wiring together TTS, ASR, LLM generation, and FFmpeg rendering end to end.
Workstation: i9-13900K · RTX 4070 Ti · 32 GB — most of what I build gets proven locally before it goes anywhere near a cloud bill.
LinkedIn · open to remote AI/ML engineering and AI-training work.