Production industrial software, on-premise AI systems, and applied research — from air-gapped IoT platforms to medical imaging.
shipped
2025 – Present
Industrial IoT Platform — Oil & Gas
Distributed platform ingesting live field-instrumentation telemetry through event-driven microservices and presenting it to instrumentation and loop-performance engineers via real-time dashboards. Integrates with plant historians (AVEVA PI, PI Vision) and field instrumentation across control loops. Decomposed a monolith into independently deployable ASP.NET Core, Node.js and NestJS services; delivered into air-gapped customer infrastructure with an offline image registry and zero-downtime rolling updates.
Air-gapped deployment • Event-driven • In production
ASP.NET CoreNode.jsNestJSKafkaMQTTKubernetesAVEVA PI
shipped
2024 – 2025
Industrial AI Co-Pilot — On-Premise RAG & Fine-Tuned LLM
Fine-tuned LLaMA 3.2 on a curated domain corpus and built a retrieval pipeline (document chunking, embedding, vector retrieval) so responses cite proprietary sources rather than model memory. Multi-step tool-calling orchestrated with LangGraph, served through FastAPI and integrated into a React frontend. Fully on-premise — no external model APIs.
On-premise • Retrieval-grounded • Fine-tuned
LLaMA 3.2RAGLangGraphFastAPIPyTorch
shipped
2024
VAE-LSTM Fault Detection
Unsupervised anomaly detection for industrial equipment, chosen because labelled failure data is scarce. A VAE learns normal behaviour across multivariate sensor streams while an LSTM models temporal dynamics; calibrated reconstruction-error thresholds surface early fault signatures. Served via FastAPI with real-time WebSocket dashboards.
Early fault detection • Multivariate sensor streams
Dual-Head Neural Network on an EfficientNetB0 backbone for simultaneous multi-class tumour classification and grade estimation from MRI. 4.0M parameters and ~15ms inference. Manuscript under review at Elsevier Results in Engineering.
93.2% accuracy • 15ms inference • Manuscript under review
Reproduction and extension of IEEE CNS 2019 work on fingerprinting encrypted voice-assistant traffic. End-to-end feature extraction and classification over network captures, with a public, reproducible repository.
Deep-learning pipeline for medical image classification across multiple imaging modalities, covering preprocessing, model training and evaluation for X-ray, CT and MRI inputs.
PythonCNNMedical ImagingDeep Learning
shipped
2024
Causal AI Anomaly Detection
Anomaly detection that applies causal inference to move from correlation-based alerts toward root-cause diagnosis for industrial processes, modelling cause-effect relationships to explain why an anomaly occurred rather than only that it did.
PythonCausal InferenceAnomaly Detection
in-progress
2025
HarmFilter App
Multi-model BiLSTM hate speech detection and prevention tool addressing harmful online content in South Asian contexts. Actively educates users to promote respectful online communities.
Multi-dataset training • Real-time detection
PythonBiLSTMNLPDeep LearningSocial Impact
in-progress
2025
FarmGuardian
AI-based decision support system for Punjab farmers. Provides crop disease detection and yield prediction through an accessible Urdu interface for illiterate and semi-literate farmers.
Real-time disease detection • Hybrid yield models
PythonDeep LearningMobile AIAgriculture
in-progress
2025
Eventure
Tourist platform for travel in Pakistan with personalized itineraries, real-time security alerts, disaster notifications, and local vendor integration.