Sarmad Khattak

Summary

What I did

• Architected complex multi-agent systems using LangChain and LangGraph, focusing on agent orchestration, inter-agent communication, and scalability.
• Built autonomous agent execution with safety rails: plan-then-execute flows, per-action
• Implemented robust RAG pipelines featuring advanced document parsing, vector retrieval, and guardrailing to ensure response accuracy and safety.
• Developed comprehensive audit logging mechanisms to monitor agent decision-making processes and ensure system transparency.
• Implemented feedback and evaluation loops: tracks user acceptance/rejection/modification of AI.
• Engineered real-time voice agents (Pipecat + WebRTC): STT → LLM → TTS with VAD and barge-in, low-latency turn handling, and Coturn/TURN credential minting (time-limited HMAC REST credentials) for reliable NAT traversal.
• Leveraged and fine-tuned Vision Language Models (VLMs) for high-level activity detection, automated alerting, and complex scene understanding.
• Engineered Person Re-Identification (ReID) systems with extensive gallery management and optimized algorithms for multi-camera tracking.
• Implemented 3D landscape reconstruction and remote sensing pipelines for geospatial analysis and environmental monitoring.
• Maintained real-time inference infrastructure using NVIDIA DeepStream and GStreamer, optimizing pipelines for state-of-the-art object detection and recognition.
• Built end-to-end voice agents for automated interviewing and customer support, integrating Speech-to-Text (STT), Voice Activity Detection (VAD), and LLM-driven logic.
• Designed low-latency conversational workflows that handle interruptions and maintain context for natural human-AI interaction.
• Developed Intelligent Document Processing (IDP) solutions capable of parsing unstructured data and performing high-accuracy OCR on diverse international formats.
• Engineered automated PII redaction systems ensuring data privacy compliance across sensitive documents.
• Integrated Generative AI for synthetic data generation to augment training sets and deployed Virtual Try-On (VTON) applications for enhanced user engagement.
• Established CI/CD pipelines using GitHub Actions, automating testing and deployment workflows across both local servers and cloud environments.
• Optimized model deployment strategies to ensure high availability and scalability of AI microservices
• Architected multi-agent AI systems, self-learning feedback loops, and agentic knowledge graphs connected to enterprise ERPs for yally.ai.
• Deployed real-time facial recognition pipelines across all international airports for IBS.
• Built automated real-time conversational voice agents using WebRTC and Pipecat for the Prime Minister Helpline.
• Engineered agentic document extraction and OCR pipelines for Ocuco.
• Deployed secure GenAI solutions and branch-level AI automation for Ubank and HBL.
• Developed the Flare platform for PTCL.

Results

• Used Arize Phoenix and Langsmith for agent observability.
• Achieved end-to-end conversational latency of under 500 milliseconds for real-time voice agents.
• Maintained sub-second query responses for high-volume enterprise document parsing and vector retrieval.