Research, measured.

62%caught, where 90% was the target

Knowing When the Object Is Not There: Calibrated Few-Shot Presence Detection for Teachable Object Recognition on Mobile Devices

MS Artificial Intelligence thesis, Bahria University

Teachable recognizers for blind and low-vision users always name an object, even when it isn’t there. In a pilot on ORBIT-India, thresholds meant to catch 90% of appearances caught 62% at best. The thesis calibrates that decision.

• #Accessibility
• #Few-shotlearning
• #Conformalprediction
• #On-deviceAI

27.45%Recall@10

How well does multilingual retrieval find Sindhi Sufi verse?

A 51-query retrieval benchmark over Shah Jo Risalo. Recall@1 9.80%, Recall@5 23.53%, Recall@10 27.45%: honest baselines on a hard, low-resource corpus.

• #RAG
• #Evaluation
• #Low-resourceNLP

92.86%precision

NeuroSymbolic AI for explainable financial fraud detection

MS AI project, Bahria University

XGBoost + SMOTE for the statistics, an OWL 2 ontology for the reasoning. 92.86% precision and a plain-language reason for every risk decision.

• #ExplainableAI
• #Neuro-symbolic
• #Imbalancedlearning
• #Ontologies

52.31%graph edges pruned

A Relevance-Aware, Multimodal Framework for Efficient Trajectory Prediction in Dense Urban Environments

Pruning 52.31% of the interaction graph before prediction, with minADE5 0.3863 m and minFDE5 0.4981 m on nuScenes. Co-authored at Bahria University.

• #Autonomousdriving
• #Trajectoryprediction
• #Efficiency
• #Graphpruning

97works mapped

Three literature syntheses: document AI, few-shot AR and agent reliability

97 works mapped across three subfields, each charted by method, result, limitation and open gap.

• #Literaturereview
• #DocumentAI
• #Agents
• #Few-shotlearning

93.33%accuracy with PSO

Training a neural network without gradients: 7 metaheuristics vs. Adam

MS AI coursework, Bahria University

Seven swarm and evolutionary algorithms train a 44-weight network head to head with Adam. Particle Swarm Optimization reaches 93.33% accuracy, 3.33 points above Adam.

• #Optimization
• #Metaheuristics
• #Neuralnetworks

92.37%test accuracy

From stroke prediction to three-tier risk stratification

MS AI coursework, Bahria University

The Kaggle stroke dataset reframed as Low, Moderate and High risk tiers. A 150-tree Random Forest reaches 92.37% test accuracy, ahead of a neural network.

• #HealthcareML
• #Imbalancedlearning
• #Modelcomparison

53%Random Forest accuracy

When deep learning loses: mental-health text classification on 992 samples

MS AI coursework, Bahria University

Anxiety, Depression, Normal or Suicidal. A TF-IDF Random Forest reaches 53% accuracy; a hybrid CNN-BiLSTM collapses to 34%. A group project about why.

• #NLP
• #Textclassification
• #Smalldata

8/8pilot invoices

Notes from a hybrid OCR + LLM invoice pipeline

Digital text first, OCR as a fallback, an LLM with a strict JSON schema on top. 8 of 8 production invoices parsed in the pilot.

• #DocumentAI
• #OCR
• #LLMextraction