Profile Photo

Yugam Padha

Machine Learning Researcher · Data Science · Explainable AI · Python · Web & API development

Email Download CV GitHub Open GitHub profile (opens in new tab) LinkedIn Open LinkedIn profile (opens in new tab) ORCID Open ORCID profile (opens in new tab)

University of Jammu — M.Sc.(Computer Science & Application) (2023–2025)

Focused on ML ,Computational theory, DBMS, Smart-Sensor, IoT, SmartWireless Systems, Operating Systems, Software engeineering along with Languages such as Python, R, Java, JavaScript, C/C++ — with a research research focussed project and real-world data modeling.

About Me

I’m Yugam Padha, an AI researcher and engineer passionate about building intelligent systems that solve real-world problems. My work spans machine learning, explainable AI (XAI), and advanced gas-sensing research, where I combine strong theoretical foundations with hands-on engineering to create models that are both powerful and interpretable.

Over the years, I’ve developed end-to-end AI pipelines for industrial gas sensing — from embedded hardware and firmware integration to data acquisition, deep learning modelling, and deployment. I also built SensorLab, a PyQt6-based desktop application that interfaces with STM32 hardware, ADS1220 ADCs, and temperature/humidity sensors, enabling real-time analytics and ML inference.

Beyond applied engineering, I actively research psychological AI, using SHAP, LIME, and hybrid ensemble architectures to create interpretable models that explore human-centered constructs such as depression, personality, and well-being. My published and ongoing research focuses on AI systems that are transparent, reliable, and grounded in domain expertise.

I’m currently preparing for a PhD in Computer Science.

I’m driven by the belief that impactful AI requires a balance of innovation, interpretability, and meaningful application. Whether advancing gas-sensing technologies or designing interpretable psychological models, my goal is to create AI systems that contribute to science, society, and industry.

If my work aligns with your vision, I’d love to connect.

Research Interests

Education

MCA – University of Jammu

Nov 2023– Oct2025 | CGPA: 8.8 (Second In Class)

Thesis: Predictive Model for Detection of Psychological Distress Using Machine Learning and Psychometric Analysis.

B.Sc. – Cluster University of Jammu

Jul 2020– Sept 2023 | CGPA: 6.4

Majors: Physics, Chemistry, Mathematics.

Publications

SVM–XGBoost Geometric–Gradient Architecture (SXGA): A Robust and Interpretable Hybrid Framework for Classification of Psychological Constructs Using DASS-21, SWLS, and BFI-10

(Manuscript submitted for peer review)

Authors: Y. Padha¹*, S. Gupta², A. Chadha³, V. Sharma⁴

This work proposes a hybrid SVM + XGBoost architecture integrating geometric margin learning and gradient-boosted reasoning to achieve superior predictive accuracy with explainable transparency.

📄 Link to Preprint

Title-page and manuscript metadata provided (under review). Future updates will include DOI or preprint link upon publication.

A Geometric–Gradient Hybrid Learning Framework for Multi-Level Explainable AI.

(Manuscript submitted for peer review)

Authors: Y. Padha¹*, A. Chadha², D. Singh³, V. Sharma⁴

This work proposes a hybrid SVM + XGBoost architecture integrating geometric margin learning and gradient-boosted reasoning to achieve superior predictive accuracy with explainable transparency.

Abstract (click to expand)

This study advances the SVM–XGBoost Geometric–Gradient Architecture (SXGA) by integrating a multi-level explainability framework tailored for psychological construct classification. Using validated psychometric instruments (DASS-21, SWLS, BFI-44), three binary outcomes—depression severity, life satisfaction, and personality adaptability—were modeled on an augmented dataset (N = 5,269). The enhanced SXGA achieved high predictive performance (F1 ≥ 0.97) while preserving feature interpretability. A tiered XAI pipeline was introduced, combining SHAP for global attribution, LIME for instance-level interpretability, Force SHAP for individualized narrative reasoning, and Counterfactuals for actionable “what-if” analysis. A licensed clinical psychologist evaluated the explanations for face validity, theoretical coherence, and trust, reporting high interpretability and alignment with clinical cognition (mean rating = 4.87/5). Notably, counterfactual analysis revealed cross-construct behavioral markers (e.g., energy and agitation), suggesting transdiagnostic patterns relevant to therapeutic intervention.

Title-page and manuscript metadata provided (under review). Future updates will include DOI or preprint link upon publication.

Ongoing / In-Progress Work

Industrial-Grade Deep Learning for Gas Sensing: Robust Calibration, Domain Adaptation & Real-Time Deployment

Status: Ongoing — data curation, model development & firmware integration

More Information (click to expand)

This paper develops an industrial-grade deep learning pipeline for concentration prediction and gas classification from multi-condition sensor streams. The work emphasizes robust calibration under temperature/humidity drift, domain adaptation across sensors and environments, and real-time inference integration with the SensorLab application and embedded STM32 pipeline.

  • Curating heterogeneous datasets across concentration ranges, gas types, temperatures and humidity; implementing systematic data augmentation and sensor noise models.
  • Designing a hybrid model (signal-processing front-end + deep regression/classifier backbone) for accurate concentration estimation and gas ID under drift.
  • Applying domain adaptation & transfer learning to generalize across sensor batches and deployment sites (unsupervised & few-shot strategies).
  • Temperature/humidity compensation modules and uncertainty-aware predictions (calibrated confidence intervals for actionable alerts).
  • Edge-ready model optimization (quantization, pruning), and integration with SensorLab (PyQt6) + STM32 firmware (CDC serial) for real-time inference and CSV/XLSX export.
  • Robustness evaluation: adversarial noise, long-term drift tests, and cross-validation across demographic/environmental stratifications where applicable.

Target outcome: a reproducible pipeline and open dataset splits plus a production-ready reference implementation that demonstrates reliability for lab → field transition.

Methods — Sparse SHAP

Sparse SHAP: Fast Approximate SHAP Calculation for Large Tabular Models

Status: Ongoing — method development & benchmarking

More Information (click to expand)

Sparse SHAP proposes a fast, approximate approach to computing SHAP-like attributions for very large tabular models by exploiting sparsity in feature interactions and using sampling + importance reweighting to reduce runtime while preserving faithfulness. The method targets production use where exact SHAP is too slow or memory-heavy.

  • Introduce sparsity-aware sampling schemes that select only a subset of feature coalitions most likely to contribute to the prediction.
  • Use importance reweighting and low-variance estimators to correct the bias introduced by sparse sampling.
  • Provide theoretical bounds on approximation error under realistic sparsity assumptions for tabular datasets.
  • Benchmark against TreeSHAP, KernelSHAP and other approximations on speed, memory, and fidelity (AUC of attribution ranking, correlation with exact SHAP).
  • Implement a PyTorch/Numpy reference implementation with CPU/GPU options, and integrate an interactive demo to visualize sparse attributions and compare with full SHAP results.

Goal: enable interpretable explanations at production scale for large tabular models without prohibitive compute cost, and provide open-source code and evaluation scripts.

*Additional manuscripts in preparation on cross-domain interpretability, human-centered AI, and counterfactual reasoning frameworks.*

Research Projects

Predictive Model for Detection of Psychological Distress Using Machine Learning and Psychometric Analysis

Original Dataset Thesis · Demo

Hybrid ML pipeline (SVM + XGBoost) trained on a self-collected psychological dataset integrating DASS-21, SWLS, and BFI-10 measures to predict affective states with explainable reasoning.

Role: Lead researcher — designed survey instruments, recruited participants, curated dataset, developed hybrid models, and evaluated interpretability.

Data Collection: Personally collected and anonymized data from participants via standardized psychological questionnaires ensuring ethical consent and balanced representation.

  • Approach: Feature engineering on psychometric scales, hybrid SVM+XGBoost architecture, and multi-level XAI pipeline.
  • Key result: Strong predictive performance and interpretable insights (detailed metrics in thesis).
  • Impact: Deployed prototype (PsychoPredictor AI) for real-time screening and clinician review.
📄 Request Thesis Copy 💻 GitHub Repository

Thesis file uploaded and referenced.

PsychoPredictor AI — Web Platform Live

Full-stack prototype enabling questionnaire input, model inference, and SHAP-driven instance explanations for end users/clinicians.

Role: Full-stack developer & integrator

Tools: FastAPI, Jinja2, Render, joblib, SHAP

  • Features: Real-time model inference, SHAP instance views, CSV batch upload, and secure input handling.
  • Deployment: Hosted on Render with joblib-serialized models and a responsive UI for clinician workflows.
  • Next: Adding Force SHAP interactive exports and clinician feedback loops.
▶ Live demo 💻 GitHub Repository (Website)

*More projects, figures, and preprints will be added as they become available.*

Ongoing Projects

In Development

Smart Gas Sensor Interface — Real-Time Data Acquisition & ML Prediction Platform Active

Developing a cross-platform app to interface with microhotplate-based gas sensors for real-time data streaming, visualization, and machine learning–driven prediction of gas concentrations under varying environmental conditions. The platform integrates hardware communication, local data storage, and cloud-ready analytics.

  • Hardware: Microhotplate array with interdigitated electrodes (IDE) for testing multiple sensing materials.
  • Core Features: Real-time sensor data logging, visual charts, temperature/humidity compensation, and export to CSV/ML pipeline.
  • Machine Learning: Embedded model inference for gas type and concentration prediction (supports SVM, Random Forest, XGBoost).
  • Cloud Integration: Planned Render or Firebase backend for remote monitoring and data synchronization.

Objective: Bridge physical sensor data and predictive ML models to enable reproducible material testing and scalable industrial gas sensing workflows.

Role: Lead developer — full-stack + firmware data handling

Stack: Python, FastAPI, SQLite, Electron (desktop UI), Matplotlib, Socket I/O

Status: Prototype phase — data streaming & visualization functional

🧠 Preview Demo (Coming Soon) ✉ Contact / Collaborate

ETA: Full desktop app release — mid 2026
(integration with gas ML models currently in progress)

Technical Skills

Programming & Development

  • Python, Java, C/C++, SQL, Git
  • API Development: FastAPI, Flask
  • Web Deployment: Render, GitHub Pages

Machine Learning & Data Science

  • scikit-learn, XGBoost, LightGBM, CatBoost
  • TensorFlow, PyTorch, Keras
  • Feature Engineering, Cross-Validation, Model Tuning
  • Smart-Sensor, AI Enabled Wireless System, Edge-AI

Data Analysis & Visualization

  • Pandas, NumPy, SciPy, Statsmodels
  • Matplotlib, Seaborn, Plotly, Altair
  • Statistical Testing & Exploratory Analysis

Deep Learning & Representation Learning

  • Neural Networks: CNNs, RNNs, Transformers
  • Transfer Learning & Embedding Interpretability
  • Experiment Tracking: MLflow, Weights & Biases (W&B)

Research & Scientific Tools

  • Jupyter Notebooks, Google Colab, VS Code
  • MATLAB (basic), R (statistical modeling)
  • LaTeX, Reproducible Research Practices

Cloud, Deployment & Infrastructure

  • Docker, GitHub Actions (CI/CD)
  • Google Cloud, AWS (EC2, S3 basics)
  • GPU Computing (CUDA environments, Colab Pro)