Abdur Rafeh B.

I design and build production-grade AI systems across computer vision, document intelligence, and LLM-based automation, taking projects from raw data to deployed, reliable pipelines. My work is built for real production environments, not proof-of-concept demos, with a strong focus on accuracy, latency, and long-term maintainability. 🔹 Core Technical Stack Languages & Frameworks: Python, C++, PyTorch, TensorFlow, OpenCV Optimization & Deployment: CUDA, TensorRT, Docker, NVIDIA Jetson Infrastructure: Linux, Git, FastAPI, REST API design Detection, Tracking & Pose: YOLO (v8+), CNNs, ByteTrack, MediaPipe 🔹 Computer Vision & Edge AI Real-time object detection and multi-object tracking on live video streams Pose estimation and body/hand/face landmark tracking using MediaPipe for applications like fitness tech, activity recognition, and gesture-based interfaces Person and player tracking with homography-based spatial mapping for sports and analytics use cases Depth sensing and 3D spatial data capture using Intel RealSense cameras Team/entity classification using clustering methods (KMeans, SigLIP-based embeddings) Model optimization for edge inference on NVIDIA Jetson and low-power hardware, cutting reliance on expensive GPU cloud infrastructure Preprocessing pipelines for degraded, low-resolution, or non-standard visual inputs 🔹 Document AI & OCR Layout-aware document parsing and structured data extraction from invoices, contracts, forms, and scanned records High-throughput OCR pipelines using OpenCV-based preprocessing to clean noisy or low-quality scans Deterministic Python validation layers around LLM-based extraction to eliminate hallucinated fields and enforce enterprise data integrity rules PDF-to-structured-data pipelines (PDF-to-SQL, database ingestion) for academic and corporate datasets 🔹 LLM, RAG & Agentic Automation Retrieval-Augmented Generation (RAG) pipelines built with LangChain, FastAPI, and the Claude API Multi-agent systems for lead generation, personalization, and data enrichment workflows Voice agents and conversational AI pipelines, combining speech-to-text (Whisper-based transcription), LLM reasoning, and text-to-speech for real-time voice interaction use cases Automated media/content monitoring systems: keyword tracking across video and text sources with automated alerting (WhatsApp, email digests) Automation workflows connecting AI outputs directly into business processes via n8n/Zapier-style integration 🔹 Data Infrastructure & Backend PostgreSQL, Supabase, MongoDB for structured and application data Distributed data ingestion pipelines from multiple sources into a unified corpus Containerized microservices (Docker) for scalable, production-grade deployment 🔹 Proven Industry Application I've applied this stack across Finance (invoice automation, audit workflows), Legal (contract intelligence, clause extraction), Manufacturing (visual inspection, defect detection), Sports/Fitness Tech (pose estimation, player tracking, analytics dashboards), and EdTech, including my current role as AI Engineer on Rubric AI, a live SaaS exam evaluation platform with paying institutional users. 🔹 How I Work Scope projects clearly before writing code, with honest, realistic feasibility assessments Communicate directly and plainly with both technical and non-technical stakeholders Build systems meant to be maintained, not throwaway scripts Stay engaged past delivery until the system performs correctly in the live production environment If you have sample data, technical specs, or an edge/voice/document AI requirement in mind, send it over and I'll tell you exactly how I'd approach it and what it would take to get it into production.