AI Engineering · Applied Research · Production Systems
Abuzar Khan
AI Systems Engineer
Building end-to-end AI systems for real-world applications
I build production AI systems using LLMs, agentic workflows, computer vision, predictive machine learning, and scalable data pipelines, from experimentation and evaluation to APIs and cloud deployment.
LLM Applications · AI Agents · Computer Vision · MLOps · Data Engineering
Based in Islamabad, Pakistan · Open to AI engineering and research opportunities
ai system lifecycle
Data Ingestion
Pipelines · ETL
Computer Vision
Detection · Segmentation
LLM & Agents
RAG · Tool use
Evaluation
Metrics · Calibration
APIs
FastAPI · Inference
Deployment
Cloud · Monitoring
data → models → evaluation → api → cloud
Overview
Professional Highlights
Engineering and research experience across the complete AI lifecycle, from data preparation and model development to deployment and production integration.
End-to-End AI Systems
Experience developing AI workflows across data ingestion, preprocessing, model development, evaluation, APIs, automation, deployment, and monitoring.
Full AI lifecycle
LLM & Agentic Workflows
Building retrieval, document intelligence, conversational AI, structured-output, tool-using, and workflow-automation systems.
Retrieval · Agents · Automation
Vision & Predictive Intelligence
Applied work involving deepfake detection, image forensics, object detection, predictive maintenance, anomaly detection, and time-series learning.
Vision · Predictive ML
Research to Production
Combining research-driven experimentation with production-oriented engineering practices for scalable and reliable AI applications.
Published AI research
What I build
Capabilities
I build AI systems across the full stack, from models and intelligent applications to APIs, automation, and deployment.
RAG and document intelligence
Conversational AI and semantic search
Tool-using agents and workflow automation
Structured outputs, evaluation, and guardrails
Python
LangChain
FastAPI
Transformers
Object detection, segmentation, and image forensics
Predictive maintenance and time-series forecasting
Fault detection and anomaly detection
Model optimization, explainability, and real-time inference
PyTorch
TensorFlow
OpenCV
Scikit-learn
ETL, data ingestion, validation, and transformation
Model-serving APIs and application integration
Cloud deployment and workflow orchestration
Monitoring, experiment tracking, and MLOps
FastAPI
Docker
Airflow
AWS
SQL
Selected Work
Selected AI Systems
Selected work across LLM applications, agentic workflows, computer vision, predictive machine learning, and production AI infrastructure.
FeaturedProduction-oriented prototype
LLM and Agentic AI
Enterprise Document Intelligence
Problem
Contracts and invoices arrive as unstructured documents that teams must read manually before any downstream system can use them.
System
An AI document-processing system that ingests documents, applies OCR and parsing, chunks and embeds the text, and uses retrieval with LLM orchestration to produce entity extractions and structured reports. Results are exposed through a FastAPI service for application integration.
My contribution
Designed the document-processing workflow, built the ingestion and preprocessing pipeline, implemented retrieval and LLM orchestration, developed structured API outputs, and created evaluation and error-analysis workflows.
Manipulated images and deepfakes are difficult to judge from a prediction alone, so a detector must also show why it flagged an image.
System
An explainable computer-vision system that detects manipulated images using deep learning and vision transformers, then produces attention-based explanation maps to localise the suspected manipulation. The work spans the DEEPFORGE final-year project and the related ViT-Xplain research on explainable detection.
My contribution
Built the image preprocessing pipeline, trained and evaluated the vision models, implemented attention visualisation for forgery localisation, and ran the error analysis used to compare model variants.
Autonomous-driving perception requires a unified spatial understanding of the scene from several camera and sensor streams.
System
A multi-camera perception prototype that preprocesses synchronised sensor streams, extracts features, builds a bird's-eye-view representation with BEVFormer, and performs 3D object detection and tracking. Built as a research prototype for autonomous-driving perception, not a vehicle deployment.
My contribution
Built the sensor data pipeline and preprocessing stage, integrated the BEV detection model, and set up the evaluation workflow used to compare detection and tracking configurations.
Industrial equipment failures surface late, when sensor drift has already turned into downtime.
System
A predictive-learning system that validates and preprocesses sensor time series, engineers temporal features, and applies residual and CRNN-style models to produce calibrated early-fault risk scores that feed maintenance alerts. It implements the direction of my published work on early-warning industrial fault detection.
My contribution
Designed the time-series preprocessing and feature pipeline, implemented and calibrated the fault-detection models, and built the evaluation pipeline used to report early-warning performance.
Lecture recordings are hard to study from when learners do not share the language of the recording.
System
A multilingual learning workflow that extracts audio from lecture videos, transcribes it with speech recognition, translates and indexes the transcript, and then answers learner questions and generates quizzes from the indexed content. Built during my TLS work on learning-content automation.
My contribution
Built the transcript and translation pipeline, implemented the retrieval and question-answering workflow, and automated quiz generation from processed lecture content.
Suspicious transactions are rare and shifting, so rule-based monitoring misses new fraud patterns.
System
A transaction-monitoring workflow that preprocesses transaction data, engineers behavioural features, and combines supervised ensemble models with anomaly detection under strong class imbalance. Scoring is exposed through an inference API for integration into application workflows.
My contribution
Built the transaction preprocessing and feature-engineering pipeline, trained and evaluated the ensemble and anomaly-detection models, and implemented the inference API.
National Centre of Artificial Intelligence (NCAI) · Peshawar, Pakistan
Supported machine-learning research through field data collection, dataset preparation, annotation, and quality validation.
Collected and curated domain-specific datasets for machine-learning research.
Conducted field visits and documented real-world data conditions and collection requirements.
Cleaned, annotated, validated, and transformed raw data into machine-learning-ready datasets.
Performed dataset-quality checks and supported exploratory model-development activities.
Maintained structured records for data provenance, annotation, and validation where applicable.
Data Collection
Annotation
Validation
Data Cleaning
Dataset Preparation
Research
Selected Research
Applied AI research across explainable computer vision, predictive maintenance, privacy-preserving learning, multi-agent systems, and intelligent infrastructure.
Featured research
ViT-Xplain: A Transparent Deepfake Detector for Consumer Electronics Based on Attention and Explainable AI
IEEE Transactions on Consumer Electronics
Published
Computer Vision and Explainable AI
An explainable deepfake-detection framework using vision-transformer representations and attention-based interpretation to improve transparency in manipulated-media classification.
Deepfake detection
Vision transformers
Explainable AI
Attention analysis
Contribution: Model development, explainability pipeline, and evaluation
IoT-Grade Robust RGB-D Small-Object Segmentation via Cross-Attention ViT Fusion and Edge Profiling
IEEE Internet of Things Journal
Published
Computer Vision and Explainable AI
An IoT-oriented small-object segmentation framework that combines deterministic data auditing, modality-specific patch embeddings, a dual-stream shared-weight Vision Transformer backbone, and quality-aware edge-guided bidirectional cross-attention for robust multimodal perception under sensor corruption and deployment constraints.
RGB-D segmentation
Vision transformers
Cross-attention fusion
Edge profiling
IoT perception
Contribution: Model architecture, cross-attention fusion design, and edge-deployment evaluation
Integrated Sensing and Edge Fusion for Consumer Vehicle Road Anomaly Monitoring
IEEE Transactions on Consumer Electronics
Published
Intelligent Transportation and Edge AI
A unified vehicle-mounted sensing and edge-computing pipeline that integrates visual, inertial and positioning streams, couples reliability-aware alignment and fusion with communication-efficient reporting, and delivers robust road-anomaly monitoring for consumer vehicles.
Integrated sensing
Edge fusion
Anomaly monitoring
Consumer electronics
Multi-modal fusion
Contribution: System architecture, multi-modal fusion design, and edge-deployment evaluation
A Survey of Dynamic Token Computation in Transformers: Taxonomy, Stability, and Budget-Aware Evaluation
IEEE Access
Published
Applied Machine Learning
A systematic survey of dynamic token computation methods in transformers, organizing approaches by taxonomy, analyzing training stability, and evaluating budget-aware inference trade-offs.
Contribution: Taxonomy, stability analysis, and budget-aware evaluation
Forensic Analysis of RGB Images Using Deep Learning - Developed a deep learning-based system to detect forged or tampered regions in RGB images, improving accuracy in identifying manipulated items for forensic analysis.
An innovative application for the citizens of Pakistan, aimed at digitalizing National Identity Cards and providing access to various government services, improving efficiency, transparency, and convenience.