Designed and deployed production-grade autonomous AI agents and workflow automations within Oracle's enterprise ecosystem (Fusion + OIC). Engineered specialized agents across Logistics, SCM, ERP, and Healthcare domains — one of the only engineers to successfully publish to the Oracle Marketplace, with agent templates released for the 26B & 26C product cycles.

Separate flagship product built at GoSaaS — transforms unstructured MD&A financial documents into structured intelligence using OpenAI Assistants. Designed the full data pipeline from raw document ingestion to PostgreSQL-backed structured output, with enterprise analytics dashboards in Oracle Analytics Cloud using custom semantic models and OTBI reports.

Adaptive Entropy-Guided Universal Domain Adaptation — a unified framework handling all four UDA scenarios (CDA / ODA / OPDA / PDA). Core contribution: dynamic entropy weighting that adapts the decision boundary based on source-target class overlap. Outperforms SOTA on standard benchmarks.

Comprehensive benchmark of FedAvg, FedSAM, FedGH, and SCAFFOLD across distributed edge nodes under non-IID distributions. Surfaces practical trade-offs between convergence, communication overhead, and privacy for real-world FL deployment.

Neural compression pipeline combining structured magnitude-based pruning with knowledge distillation from a teacher model. Achieves significant parameter reduction while preserving task accuracy — targeting mobile and edge deployment.

Multi-label infrastructure damage detection benchmarking ResNet50, InceptionV3, and VGG16. Custom loss weighting strategy to address severe class imbalance in real-world road survey datasets.

Multi-agent GenAI framework for co-creative children's storytelling. Integrates LLM narrative generation, TTS audio narration, and text-to-video synthesis in a single pipeline — with Urdu language support and child-safe content moderation. Developed as BSc capstone at LUMS and published through CSaLT.

Full-stack F1 intelligence platform — XGBoost race predictor, Monte Carlo tyre simulator, Llama 3.2 AI analyst, lap-by-lap telemetry, historical constructor data (1950–present), and automated session email reminders.

RAG-powered legal document analyzer on Mistral 7B. Ingests statutes and contracts into a FAISS vector store; answers natural-language queries with source attribution.

LLM-powered academic planning engine. Ingests course syllabi and student profiles; generates personalised academic roadmaps using Retrieval-Augmented Generation via LangChain.

Reconstructs 3D landmark models from 2D image sets using ORB feature detection, triangulation, bundle adjustment, and mesh generation. Interactive 3D visualization via FlutterCube. Bridges classical CV with modern rendering pipelines.

Full-featured HR self-service portal — attendance tracking, leave management, payroll access, and automated email workflows with role-based access control for employees and administrators.

Real-time B2B trading platform with WebSocket-powered live competing offers, item bartering and cash transactions, and full trader profile management.

Animated restaurant website built for a real client — scroll-driven story section, interactive menu with pricing card animations, and a layered CSS micro-animation system.

Next-generation interactive portfolio featuring a Three.js Galaxy Explorer, real-time System Monitor, and in-page LLM Chat — built with Next.js, Three.js, and Framer Motion.

Kahaani: A Multimodal Co-Creative Storytelling System EACL SRW 2026 — Accepted & Published Featured by University of Warwick Education Fund