Yousif Alsaffar's Work | ContraWork by Yousif Alsaffar
Yousif Alsaffar
pro

Yousif Alsaffar

Agentic AI Engineer | Cloud, Backend & Low-Latency Systems

    2
    Followers
Cover image for From September to December 2024
From September to December 2024 I was a Software Engineer Fellow in the MLH Fellowship, contributing to Meta Research's AEPsych: an open-source framework for adaptive experimentation in psychophysics and perception research, built on GPyTorch and BoTorch. What I did: applied Gaussian Process models (PyTorch, BoTorch, CUDA) so human-perception studies reach the same answer with fewer trials, working directly with Meta researchers and maintainers and shipping changes through their open-source review process. I was later named one of MLH's Top 50 Hackers of 2025.
0
19
Cover image for S3TVR (Seamless Speech-to-Speech Translation with
S3TVR (Seamless Speech-to-Speech Translation with Voice Replication) was my bachelor's thesis at IE University. It translates live speech between English and Spanish and speaks the translation in the speaker's own voice, cloned zero-shot, running locally. How it works: a cascaded real-time pipeline of voice activity detection, noise reduction, speech recognition (NVIDIA Parakeet / FastConformer via NeMo), translation (Meta NLLB-200) and voice-cloning text-to-speech (Coqui XTTS v2), optimised with CUDA and DeepSpeed. Every stage is modular, so models can be swapped as better ones appear. Results against Meta's Seamless Streaming model: about 3.1 s average latency per sentence versus 4.6-4.7 s, and BLEU of 0.36 (English) and 0.41 (Spanish) versus about 0.20.
0
24
Cover image for OpenIndexAI is an open-source desktop
OpenIndexAI is an open-source desktop app that turns everything on your computer into a searchable visual gallery. All AI runs locally on your GPU; the only network call is a one-time model download. What I built: fast incremental indexing (parallel directory walk, persistent SQLite index, rescans that only touch changed files); on-device AI with ONNX Runtime + DirectML for CLIP semantic search, face detection and clustering into people, and OCR text search; hybrid search that fuses OCR text, filenames and CLIP vector similarity; and safe bulk actions, with collision-safe moves and a three-stage delete that stays reversible at every step. A swappable model layer lets new local or API-backed models plug in without rewriting the app. Built with Rust, Tauri and TypeScript. Early alpha on Windows, images today, with video and documents next.
0
26
Cover image for Wahid.AI (https://Wahid.AI) is an Arabic
Wahid.AI (https://Wahid.AI) is an Arabic AI study companion for Iraqi secondary students preparing for the ministerial exams that decide their university future. I founded it and built the full stack. What it does: Zaid, an AI tutor, answers questions about any page of the official textbook. Students practise on 8,154 real ministerial exam questions, and AI grading marks typed, handwritten (photographed) or spoken answers instantly and explains where marks were lost. Plus flashcards, the full textbooks, and nearly 20,000 lessons from 111 Iraqi teachers, organised chapter by chapter. How it's built: a multi-step Arabic RAG system over a vector database powers the streaming tutor, semantic search and grounded grading. The backend is an ETL pipeline and multi-database stack on Google Cloud with autoscaling, caching and model routing, running at about $0.05 per active student per month. Results: 21,957 students (78% in their final exam year), 38,678 practice exams, 56,066 answers graded by AI and 19,674 questions asked to Zaid. When students retry a question they got wrong, their correct-answer rate rises from 40.5% to 54.8%. Free for students; schools pay for a dashboard showing each student's weak chapters and exam readiness.
0
31