Huzaifa Naveed

Built a 3D face reconstruction pipeline using 3D Gaussian Splatting with FLAME mesh initialization via MICA and PnP-based pose estimation for photorealistic novel view synthesis. Implemented a thumbprints and signatures detection pipeline using YOLO26 combined with handcrafted rules. Fine-tuned SmolVLM on a custom dataset to create a robust bilingual English-Urdu OCR extractor.

Deep Learning Engineer at Unikrew Solutions (2026-06 – Present)

• Built a 3D face reconstruction pipeline using 3D Gaussian Splatting with FLAME mesh initialization via MICA and PnP-based pose estimation for photorealistic novel view synthesis.
• Implemented a thumbprints and signatures detection pipeline using YOLO26 combined with handcrafted rules.
• Fine-tuned SmolVLM on a custom dataset to create a robust bilingual English-Urdu OCR extractor.

AI Trainee at Unikrew Solutions (2025-09 – 2026-05)

• Designed and trained a Multiscale Attention Fusion Network for deepfake detection, achieving 97.45% accuracy across StyleGAN, Diffusion, and real sources, surpassing SOTA models.
• Extended the transliteration pipeline with gazetteer lookup and fallback logic, surpassing ChatGPT on accuracy and consistency.
• Built a CNIC verification pipeline validating documents against security features, flagging tampered submissions.
• Designed a custom Urdu OCR architecture combining a Vision Transformer, dilated Conv1D projection, and CTC loss for character prediction.

AI Intern at Unikrew Solutions (2025-07 – 2025-08)

• Fine-tuned a quantized LLaMA 3.1 8B model using LoRA on a custom dataset for Urdu-to-Roman transliteration.
• Contributed to a U²-Net based document integrity model to detect manipulation in scanned documents.
• Fine-tuned Whisper on a labeled Urdu speech dataset to improve Urdu speech-to-text performance.

Bachelor of Science in Artificial Intelligence – FAST NUCES Karachi (2022-08 – 2026-06)