Akthar N.

With 4+ years in AI, I bridge the gap between advanced research and production engineering. I don't just train models; I build the scalable infrastructure to serve them. My expertise takes custom neural networks from mathematical theory to fully deployed, production-grade APIs on Kubernetes clusters. 🛠️ Core Technical Stack 1. Generative AI & Agents Tech: LLMs, AI Agents, RAG, LiveKit, TTS, STT, LangGraph, MCP, n8n workflows. Memory & Context: mem0 (Memory Management), RAG pipeline optimization. 2. Deep Learning & Neural Networks Core: PyTorch, Stable Diffusion, Transformer, 3D U-Net, IP-Adapter, ControlNet, ArcFace. 3. Backend API Development & Real-Time Streaming Frameworks: Expert in FastAPI (Async/High-concurrency) and Flask for serving ML models. Protocols: RESTful APIs, WebSockets, gRPC, and Swagger/OpenAPI documentation. Streaming: LiveKit, WebRTC, Socket.io for low-latency interactive AI. 4. MLOps, DevOps & Big Data CI/CD & Monitoring: GitHub Actions, Sentry (Error Monitoring), Automated Testing Pipelines. Infrastructure: Docker, Kubernetes (Scalable Clusters). Data Engineering: Kafka, Debezium, PySpark for real-time data streaming. HPC: Distributed training (Multi-node/Multi-GPU) on A100 & H100 clusters. 📂 Project Showcase 🤖 Intelligent Voice Agent & RAG System Role: Led the upgrade of a Voice Agent (v0.x to v1.x), serving the RAG pipeline via optimized FastAPI endpoints. Agentic Workflow: Implemented long-term memory using mem0 and integrated n8n workflows via an MCP server. Ops: Deployed on Kubernetes with GitHub Actions for CI/CD and Sentry for real-time error tracking. 🎥 Real-Time Talking Avatar Role: Integrated LiveKit (WebRTC) with Wav2Lip for a live-streaming avatar solution. Performance: Achieved sub-200ms latency on GPU clusters using optimized socket streaming. Impact: Reduced deployment costs by 50% through rigorous model pruning, quantization, and fine-tuning. 🚀 Stable Video Diffusion & HPC Training Role: Developed distributed training code for complex architectures (3D U-Net, IP-Adapter, ControlNet) using DensePose and 3D pose datasets. Infrastructure: Trained on 4x A100 (single node) and 2x H100 (multi-node) clusters. Result: Achieved 40% faster convergence through multi-node optimization compared to single-GPU training. 🗳️ Election Prediction System (Oman Shura Election) Role: Engineered a real-time election analysis pipeline using Debezium, Kafka, and PySpark. Modeling: Deployed Transformer-based models on 4x A10 GPUs to forecast polling percentages for 63 regions. Visualization: Created interactive dashboards visualizing 6 years of historical voting data. 💬 Speech-to-Text (Whisper ASR) Role: Trained and deployed OpenAI’s Whisper model using the Libri-speech dataset. Impact: Created a high-availability API pipeline capable of handling diverse audio inputs.