Uzair M.

My production systems are used by millions of people every day. Over 40 banks and financial institutions across multiple countries rely on AI infrastructure I built — from passport validation and deepfake detection to multilingual OCR, agentic RAG, and facial liveness verification. What separates my work from most ML engineers: I do not stop at the model. I build the APIs, the backend infrastructure, the mobile integrations, and the edge deployment pipelines that make AI actually run in production — on constrained hardware, at scale, without cloud dependency. If you need it to work in the real world, not just in a notebook — that is what I do. ✅ What I Offer: 🔹 End-to-End AI System Development From model design to full production deployment — APIs, backend systems, mobile integration, and edge-ready pipelines 🔹 LLMs, Generative AI & RAG Systems Fine-tuning (LoRA, QLoRA, SFT, DPO), agentic workflows, RAG pipelines, multilingual AI, and voice-based intelligent agents 🔹 Computer Vision & Document AI Facial recognition, liveness detection, passport validation, MRZ extraction, OCR (including low-resource languages), ICAO-compliant processing 🔹 Deepfake Detection & Fraud Prevention Production-grade anti-spoofing systems with ultra-low false accept rates for secure eKYC and identity verification 🔹 AI for Fintech & eKYC Solutions Complete onboarding pipelines used by banks — document verification, facial matching, fraud detection, and compliance-ready systems 🔹 Data Annotation & Dataset Engineering High-quality image and video annotation (bounding boxes, segmentation, keypoints, tracking) Custom dataset creation, cleaning, and labeling pipelines optimized for model performance ------------------------------ 🎯PRODUCTION METRICS — LIVE SYSTEMS Every number below is from a system running in production right now: 🔹 RAG-based voice agent: under 5 sec end-to-end latency | Precision@K above 98% 🔹 Gen AI transliteration agent: under 1% Word Error Rate 🔹 Facial Recognition system: under 1% FRR | Deployed on CPU at scale with less than 5 sec latency 🔹 Document & passport validation: under 1% false rejection rate 🔹 Deepfake detection: 0.3% false accept rate | zero fakes pass | beats published SOTA 🔹 Custom OCR (Urdu & low-resource): under 1% Word Error Rate 🔹 ICAO background removal: production-certified | live in eKYC flows These are not benchmarks. These are live metrics on systems processing real user data at this moment. ------------------------------ ✅ CORE EXPERTISE 🔹 LLM Fine-tuning & Generative AI I fine-tune Llama and GPT-family models for domain-specific tasks — transliteration, document-grounded generation, structured extraction. Full pipeline: LoRA, QLoRA, SFT, and DPO alignment. 🔹 Computer Vision & Document AI Passport validation, CNIC processing, MRZ extraction, ICAO-compliant background removal, facial comparison, occlusion detection, and liveness verification. My masked facial recognition system using diffusion models was developed as graduate thesis research. 🔹 Deepfake & Fraud Detection Deployed deepfake detection that surpasses current published SOTA — 0.3% false-accept rate, zero fake approvals in production. One of the most reliable anti-spoofing systems in active eKYC deployment today. 🔹 AI Agents & RAG Systems Production-grade agentic pipelines — RAG-based voice agents, tool-calling systems, multi-step reasoning workflows. Sub-5-second latency, Precision@K above 98%, running live. 🔹 eKYC & Fintech AI Infrastructure End-to-end AI backbone for digital onboarding systems trusted by 40+ banks — decency verification, facial liveness, document fraud detection, multilingual OCR. From Model to Production — APIs, Mobile & Edge This is where most ML engineers stop. I do not. I build: 🔹 Scalable REST APIs and ML serving pipelines (Python, FastAPI, chi, Golang) 🔹 Concurrent, high-throughput backend systems in Golang for low-resource and CPU-only environments 🔹 Android mobile applications in Kotlin with on-device ML integration 🔹 Edge deployment pipelines — quantized, optimized models running on CPU at scale in production If a model needs to run on a bank's on-premise server, a mobile device, or an edge node with no GPU — I have already done it. ------------------------------ ✅ TECHNOLOGIES 🔹 Vision: YOLOv8/v9, DeepLabV3+, Segformer, ViT, Swin Transformer, OpenCV, MediaPipe 🔹 LLMs & NLP: Llama 3.x, GPT-4, BERT, HuggingFace Transformers, LangChain, PEFT (LoRA/QLoRA) 🔹 ML Frameworks: PyTorch, TensorFlow, Keras, MMDetection, MMSegmentation 🔹 Backend: Python, Golang, FastAPI, REST APIs, concurrent systems 🔹 Mobile: Kotlin, Android, Jetpack Compose, on-device ML 🔹 Deployment: Docker, edge deployment, CPU inference, Streamlit, Colab 🔹 Data & Infra: NumPy, Pandas, Scikit-learn, Matplotlib ------------------------------ 🎙️LET'S TALK If you are building an AI solution that is scalable to a wider audience, I can add serious value to the end product. Feel free to drop a message