Farhan Sheth
Scholar viewDownload CV

Navigation

Download CV

Research case studies

Projects

I turn research questions into datasets, models, evaluation pipelines, and usable tools.

The projects below connect applied AI methods with measurable outcomes and public artifacts.

Case study

RecruitView & CRMF

A multimodal interview dataset and manifold-fusion framework for personality and interview-performance assessment.

RecruitView treats interview assessment as a data and representation-learning problem, connecting naturalistic video collection with comparative human judgements and a lightweight multimodal model rather than presenting a generic screening interface.

Problem
Assess personality and interview performance from naturalistic multimodal interview data without reducing the task to a single modality.
Role
Project lead for the grant-supported interview-assessment collaboration; led the RecruitView data-collection and annotation infrastructure and designed CRMF.
Research contribution
Built the end-to-end data and modelling pipeline, including the psychologist-informed QA-Labeler platform and geometry-aware Cross-Modal Regression with Manifold Fusion.
Dataset scale
  • 2,011 naturalistic video interview clips
  • More than 300 participants
  • 27,000 pairwise comparative judgements across 12 dimensions
Methodology
  • Video, audio, and text embeddings
  • Manifold-specific experts over hyperbolic, spherical, and Euclidean spaces
  • Adaptive routing and pairwise comparative annotation
Measured outcomes
  • Up to 11.4% higher Spearman correlation
  • Up to 6.0% higher concordance index
  • 40–50% fewer trainable parameters than large multimodal baselines
Technology
  • PyTorch
  • Transformers
  • NLP
  • Web APIs
  • MongoDB

Case study

HERALD: Privacy-Preserving Clinical LLM

A model-agnostic, client-side cryptographic redaction framework for clinical language-model deployment.

The case study focuses on deployable privacy-utility trade-offs for cloud and API-assisted clinical NLP: selected tokens remain protected while the surrounding context stays usable for evaluation and real workflows.

Problem
Protect selected sensitive content in clinical LLM workflows while retaining the contextual information needed for downstream utility.
Role
Research Assistant at NTU; developed and evaluated the framework with Dr. Si Yong Yeo.
Research contribution
Designed a model-agnostic client-side pipeline that applies medical NER and part-of-speech policies before deterministic ciphertext replacement.
Dataset scale
  • Public clinical classification workloads
  • Public medical question-answering workloads
Methodology
  • Medical named-entity recognition
  • Part-of-speech policy selection
  • Deterministic client-side ciphertext replacement
Measured outcomes
  • Preserves protected tokens through storage, transmission, and model processing
  • Recovers substantially more downstream utility than fully secured baselines
  • Requires no downstream model architecture change
Technology
  • PyTorch
  • Transformers
  • NLP
  • Cryptography
  • Large language models

Case study

Agro Companion

An agricultural decision-support tool combining soil-image classification with fuzzy crop recommendations.

Agro Companion turns image classification into an agricultural decision path: the soil prediction feeds a transparent recommendation layer instead of producing an isolated label.

Problem
Help farmers connect soil-image evidence with crop recommendations that also account for geological and environmental factors.
Role
Led the research framework and translated the soil-classification work into a public decision-support application.
Research contribution
Combined CNN and vision-transformer soil classification with fuzzy crop recommendation rather than treating recognition as the final decision.
Dataset scale
  • Seven soil types
  • More than 20 crop recommendations
  • Dataset expanded from 1,189 to 8,413 images
Methodology
  • CycleGAN augmentation
  • CNN and ViT families with grid search and k-fold validation
  • Ensemble learning with fuzzy logic
Measured outcomes
  • 7.1× dataset expansion (+607.6%)
  • 92% crop-recommendation top-1 accuracy
  • Recommendations combine five geological and four environmental factors
Technology
  • Python
  • PyTorch
  • Keras
  • OpenCV
  • Flask
  • MATLAB
  • Fuzzy logic

Case study

Herbify

A curated medicinal-herb image dataset, ensemble vision framework, and real-time identification application.

Herbify connects curation, preprocessing, model evaluation, and public identification in one evidence trail without replacing species-level uncertainty with a decorative prediction surface.

Problem
Make medicinal-herb identification more reliable through a cleaned dataset, explicit preprocessing, and an accessible real-time application.
Role
Curated the dataset, developed the ensemble vision framework, and translated the work into the public identification application.
Research contribution
Standardized the Herbify dataset with PAHD preprocessing and paired EfficientNetV2-Large with ViT-Large/16 for species identification.
Dataset scale
  • 6,104 curated images
  • 91 herb species
Methodology
  • PAHD preprocessing and quality control
  • CNN and vision-transformer fine-tuning
  • EfficientNetV2-Large and ViT-Large/16 ensemble
Measured outcomes
  • 99.56% F1 score
  • Scientific-name and resemblance-probability output
  • Accessible real-time identification application
Technology
  • PyTorch
  • OpenCV
  • Scikit-Learn
  • Flask
  • CNNs
  • Vision Transformers