Ahmad Zafar Khan

Lead AI Engineer | Enterprise ML Systems & LLM Optimization Expert

Based in: Lahore, PakistanMost recently: Lead AI Engineer, MW Apps GmbHExperience: 4 years

Lead AI Engineer with a track record of architecting high-performance ML systems for enterprise scale. Expert in optimizing model lifecycles—from RAG architectures with LangChain to low-latency inference using TensorRT and quantization. Proven ability to bridge the gap between complex AI research and production-ready business solutions.

Experience

Lead AI Engineer

MW Apps GmbHMonheim am Rhein, Germany (Remote)work2025

Summary

Leading end-to-end development of ML systems for enterprise applications, ensuring optimized performance and low-latency inference. Managing model lifecycle from data cleaning and feature extraction to production deployment using FastAPI and TensorRT. Conducting evaluation, benchmarking, and model selection for large-scale NLP and vision-based systems. Collaborating with cross-functional teams to align business objectives with ML deliverables and document technical workflows.

Results

Optimized model inference and deployment pipelines, reducing resource footprint through quantization and inference optimization.
Scaled system to handle production workloads without model bottlenecks, ensuring enterprise-ready stability and predictable resource usage.
Achieved system predictability under sustained production load, shifting the model from a prototype bottleneck to a reliable enterprise component.
Reduced latency and improved system throughput for higher request volumes.
Eliminated model inference as the primary system bottleneck through quantization and optimization.

FastAPITensorRTNLPComputer VisionML Systems

AI/ML Engineer

Verdant SoftLahore, PakistanworkApr 2024 – Dec 2024

Summary

Fine-tuned LLMs (GPT, BERT, Falcon) using LoRA and PEFT, improving contextual accuracy by 30%. Developed Retrieval-Augmented Generation (RAG) systems with FAISS and LangChain, supporting enterprise knowledge-grounded assistants. Built modular ML data pipelines for preprocessing, model training, evaluation, and inference optimization. Applied quantization (GPTQ, ONNX, Llama.cpp) and TensorRT acceleration to achieve 40% faster inference. Implemented RLHF (Reinforcement Learning from Human Feedback) for conversational model alignment. Collaborated with product managers to integrate ML logic into business workflows and create documentation for production models.

Results

Built scalable ingestion and preprocessing pipelines for large text and multimodal datasets using PySpark on Databricks clusters, eliminating single-node memory bottlenecks.
Parallelized schema validation, deduplication, and metadata normalization to ensure high-quality data for downstream LLM fine-tuning.
Optimized the data lifecycle by staging raw payloads in S3 and utilizing Parquet and Delta formats for efficient access and storage.
Improved contextual accuracy by 30% through LLM fine-tuning and achieved 40% faster inference using quantization and TensorRT acceleration.

GPTBERTFalconLoRAPEFTRAGFAISSLangChainGPTQONNXLlama.cppTensorRTRLHFPythonAWSDatabricksPySparkS3Delta Lake

AI/ML Engineer

CacheLogicLahore, PakistanworkNov 2022 – Mar 2024

Summary

Designed and deployed an AI-powered real-time sound classification system for mobile platforms, integrating time-series analysis and signal-based ML models. Built face and device detection pipelines using CNNs for privacy-preserving media applications. Created large custom datasets for supervised learning tasks (object detection, image segmentation). Implemented diffusion models and GANs for user-personalized animated character generation. Developed and deployed mobile-integrated ML APIs using Flask and PyTorch. Optimized ML workflows for real-time inference and memory efficiency on edge devices.

Results

Managed and iteratively refined custom datasets ranging from 10,000 to 100,000 images, focusing on diversity and annotation accuracy to resolve model failure cases.
Developed and deployed real-time sound classification and vision pipelines (face/device detection) optimized for edge device memory efficiency.
Scaled datasets up to 100,000 images to achieve stable production performance for detection and segmentation models.

CNNsDiffusion ModelsGANsFlaskPyTorchTime-series analysisObject DetectionImage Segmentation