22 Research Papers Across Three Sessions: What Researchers Actually Presented at PiCET 2026

Three technical paper presentation sessions across the two days of PiCET 2026 hosted twenty-two research papers. The day-one session at CV Raman-302 was evaluated by Mr. Chirag Patel, external evaluator from CHARUSAT, alongside Dr. Kamal Sutharia, Head of the AI/ML Department at Parul University. The day-two paper presentation and pitching session was evaluated by senior faculty across academic and industry backgrounds. The day-two morning session in the CV Raman Building was judged by Dr. Abhilash Patil as the external evaluator and Mr. Pravin Kumar Patidar as an associate professor at Parul University, in a hybrid format combining offline and online presentations.

The papers spanned the full breadth of contemporary applied machine learning. They ranged from doctoral research with published-paper depth to early-stage undergraduate explorations. They included papers from across India and from an international researcher joining online. They covered healthcare emergencies, civilian protection in conflict zones, accessibility for users with dyslexia, the prediction of river floods, and the democratisation of machine learning for non-technical domain experts. Organising the highlights by research theme produces a clearer view of where contemporary applied AI is moving than session order alone.

Three healthcare-focused papers across the sessions addressed conditions where early or accurate detection has measurable consequences for patient outcomes.

Ms. Neesha Kumari Raye presented a hybrid machine learning approach for predicting postpartum depression risk. The architecture combines an autoencoder for feature compression and noise reduction with an XGBoost classifier for the final risk categorisation. Input data came from structured health questionnaires. The system achieved accuracy above 96 percent across low, moderate, and high risk classifications. Correct identification of high-risk individuals is the critical clinical metric in this domain. The evaluator noted that data bias, ethical considerations, and generalisability of results across populations remain open research questions that future work should address.

Ms. Divya Hasmukhlal Panchal presented work on glaucoma detection that integrates structured patient data with optical coherence tomography (OCT) imaging data through a hybrid ML framework. The distinguishing feature of the approach is its time-awareness. Rather than treating glaucoma diagnosis as a single classification problem at one timepoint, the model accounts for disease progression dynamics and makes longitudinal predictions about disease development. The integration of multiple data sources produces a deeper diagnostic picture than imaging-only or numerical-only approaches. The use of multimodal and longitudinal datasets for complex healthcare problems represents a recent direction in clinical AI research.

Ms. B. Deepthi presented work on brain MRI classification for tumour detection using a hybrid ResNet-152 framework with global-local feature fusion. The model analyses each MRI image at two scales simultaneously: extracting fine local details while preserving global brain-structure context. Image inputs are standardised at 224 by 224 pixel resolution. The initial layers of the ResNet-152 model are frozen to retain learned representations of basic shapes and textures, while subsequent layers are trained on brain MRI specifically. The data is split 80-20 between training and testing. Brain tumours vary significantly in shape, size, and colour, making accurate detection a structurally difficult problem that conventional single-scale models struggle with.

Also Read: Reimagining Universities in the Age of AI With Mr Aravind Voruganti

Two agriculture-focused papers addressed the application of machine learning to two distinct agricultural problems.

Mr. Chirag Tiwari presented an image-based deep learning system for detecting diseases in tomato plants using the Xception architecture with transfer learning. The multi-class classifier can detect bacterial spot, early blight, leaf mould, and other tomato leaf diseases. The distinguishing feature is the integration of Grad-CAM for explaining the model’s decisions, allowing users to verify whether the model is focusing on the correct parts of the leaf for diagnosis. The system shows high classification accuracy under controlled conditions. Real-world agricultural deployment requires additional work given the variability in image quality from field-deployed cameras. Students entering the B.Tech in Biotechnology at Parul University study the agricultural application layer that papers like this address.

Ms. Saba Anjum Jahangir Patel, faculty at Vishwakarma University, presented a modular Retrieval-Augmented Generation (RAG) architecture for delivering smart farming advice to small farmers. The system processes soil fertility, current weather conditions, and crop type simultaneously, then uses large language models including Llama and Mistral to generate context-specific recommendations. The application addresses two practical barriers to AI adoption in Indian agriculture: the lack of local-language interfaces (the system supports Marathi and Hindi alongside English) and the absence of explanation for AI recommendations (the system provides traceable sources for each piece of advice). Farmer trust depends on knowing where advice came from before acting on it in the field.

Three papers addressed the application of AI to educational contexts, from student performance prediction to classroom assistance to language-specific question answering.

Mr. Kartik Gori presented work on the trade-off between predictive accuracy and model interpretability in educational AI. Traditional machine learning models function as black boxes, which limits their utility in education where teachers and administrators need to understand why a student is predicted to underperform. The study used approximately 14,000 student records with variables including attendance rate, daily study time, completed assignments, and socio-demographic characteristics. Multiple models were compared: Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, and Neural Networks. The principal contribution is the integration of SHAP (SHapley Additive Explanations) for both global and local explanation of predictions. SHAP can identify specific determinants behind a poor-performance prediction (low attendance, high stress, irregular study time), allowing educators to intervene on actual causes rather than generic risk profiles.

Mr. Ashishkumar Patel, faculty at LDRP Gujarat, presented a conceptual approach for integrating AI chatbot tools into classroom learning. The motivation is the dual pressure on educators: too much classroom emphasis on textbook learning at the expense of project-based skill development, and the structural impossibility for a single teacher to attend to every student in a large classroom. AI chatbot integration allows students to ask questions throughout the day and night, including questions they might be embarrassed to ask in front of peers. The paper presented an idea-stage design rather than an implemented system. The clarification from the evaluator was that AI is not replacing teachers but is removing the routine knowledge-delivery load, freeing teachers to mentor students on actual project work.

Mr. Baldha Niravkumar Amrutlal, research scholar at Parul University, presented a survey on extractive question answering using machine reading comprehension for the Gujarati language. The motivation is the structural bias of contemporary natural language processing tools toward English, French, and Chinese, with limited support for Indian languages. The Gujarati language presents specific computational challenges: agglutinative morphology where word suffixes change with usage, and relatively free word order within sentences that preserves meaning despite re-arrangement. The proposed direction is the construction of a clean Gujarati dataset for training models to process Gujarati natively rather than through translation. The conversion-to-English approach introduces noise that degrades performance. The next-step framework includes TensorFlow as the implementation environment.

Two papers addressed the recommender system domain, including the only student-led paper presented offline (in person) on day two. Mr. Mohmad Ali Unawala, an undergraduate student at the Parul Institute of Information Technology, presented his outfit recommender work to a panel of faculty and external evaluators while the other presenters of the session joined via video conferencing. The choice to present in person rather than online was, in his case, a structural test of student readiness for direct academic evaluation.

Check Out: Faculty of Engineering and Technology at Parul University, programs that prepare you for all the new tracks of research.

Mr. Mohmad Ali Unawala identified a real limitation of conventional online shopping recommendation systems: they recommend similar items (more shirts when a shirt is searched) rather than compatible items (pants that go with the shirt). The proposed solution uses Graph Neural Networks (GNNs), where clothing items are represented as nodes and compatibility relationships are represented as edges. The training dataset was Deep Fashion Multimodal, with over 40,000 colour images of garments and 12,000 mapping images across 17 clothing categories. The system enforces category coherence (men’s shirts paired with men’s pants, not women’s). The model achieved a score of 0.665, substantially outperforming previous GNN baselines that scored 0.08. The implementation was built using Kaggle with Python.

Mr. Ashishkumar Patel presented a hybrid approach to recommender systems combining multiple similarity metrics through mathematical amalgamation rather than selecting a single similarity measure. The test dataset was the MovieLens benchmark. The amalgamation approach combined user-similarity and item-similarity equations to produce richer recommendations than either measure alone. The new hybrid approach demonstrated 10 to 12 percent improvement over single-model baselines. The implementation used Java and Python.

Two papers addressed core computer vision problems at different scales and with different applications.

Mr. Harshil Narendrabhai Chauhan, PhD scholar at the Department of Computer Science and Engineering at Parul University, presented work on image captioning guided by Dr. Chintan Bhupeshbhai Thacker, Associate Professor. The architecture, called Gated Feature Fusion Image Captioning Model, uses two visual encoders in parallel: Xception for global semantic understanding and YOLOv8 for object detection with bounding boxes. The contribution is a learnable Gating Mechanism that, at every spatial location, computes a weight between 0 and 1 to determine how much to rely on global semantic features versus object-level features. The decoder uses Bahdanau Attention with a Gated Recurrent Unit (GRU) to generate the caption word by word. On the Flickr8k benchmark, the model achieved BLEU@1 of 0.774, BLEU@2 of 0.658, BLEU@3 of 0.462, BLEU@4 of 0.371, and METEOR of 0.452, outperforming seven baseline models. Ablation studies confirmed that removing the gating mechanism degraded performance, validating the architectural contribution. The Ph.D. in Engineering and Technology at Parul University is the academic home of this research.

Ms. Jimbiwa Kajimalwendo from the Department of Mechatronics Engineering at PIET, Parul University, with co-authors Dr. Prabodh Kumar Sahoo and Dr. Heli Shah, presented work on detecting moving objects in satellite video footage. The problem is structurally hard because the camera itself moves (the satellite orbits), objects are often only a few pixels in size, frame rates are low, and environmental conditions including cloud cover and lighting variation introduce noise that mimics motion. The proposed Temporal Differencing-based Background Modelling (TDBBM) method builds an initial background model from the first 100 frames, uses three-frame differencing to identify moving pixels, and updates the background continuously through an Infinite Impulse Response (IIR) filter. An advanced extension called AMEP-GMS adds multi-spectral phase-gradient motion segmentation. A further HiEUM framework achieved an F1-score of 89.7 percent at 98.8 frames per second, making the system viable for real-time satellite surveillance.

Three papers addressed contemporary security and safety problems with measurable real-world consequences.

Mr. Bilalkhan R. Pathan presented work led by a seven-member team (Syed Ibad Ali, Bilalkhan R. Pathan, Himadri Vegad, Sapna V. Bhimajiyani, Karnavi Desai, Himani Parmar, and Ziyam Khan) from the AIDS, CSE, and IT departments at Parul Institute of Engineering and Technology. The paper addresses civilian evacuation during active missile conflict, a problem that combines machine learning, geospatial intelligence, reinforcement learning, and systems architecture. The AI Processing Engine integrates satellite imagery, real-time threat alerts, population density, infrastructure data, road network conditions, and shelter capacity to produce dynamic risk heatmaps and adaptive evacuation routes. The four-module pipeline covers Risk Assessment, Safe Shelter Identification, Adaptive Routing via reinforcement learning, and Load Balancing across shelters. Compared against shortest-path routing and static shelter assignment baselines, the proposed system reduced average evacuation time from 32.4 minutes and 27.9 minutes to 18.7 minutes, increased successful evacuation rates from 74.1 percent and 81.2 percent to 93.6 percent, and reduced shelter congestion from 61.7 percent and 49.8 percent to 24.3 percent. The system includes end-to-end encrypted communications and blockchain-based event logging for post-conflict accountability.

Ms. Garima Sharma, faculty at Poornima University, presented work on Split Learning for wearable health devices. The structural challenge is that small wearable devices collect substantial personal health data but lack the computational capacity to process it locally, while sending raw data to remote servers raises privacy concerns. The SAPSL approach splits the neural network: a small, secure portion runs on the wearable device itself and encrypts/scrambles sensitive data before transmission, while the heavier computation runs on the main server. A reinforcement learning component dynamically adjusts the split point based on available battery. The system achieved approximately 89 percent accuracy across testing with 10,000 samples and 60,000 training samples. The evaluator noted that the 89 percent figure reflects an inherent trade-off: stronger privacy protection through data scrambling sacrifices a measurable fraction of accuracy. For highly secured wearable systems, 89 percent is an excellent operating point.

Mr. Akash R Jadhav from PES University, Bangalore, presented work on detecting deepfake and AI-manipulated images that combines federated learning (for privacy) with centralised learning (for accuracy). A small client-side application of approximately 10 megabytes runs in the user’s browser and performs a fast preliminary verification without sharing the image externally. Only if the local model cannot decide with high confidence does the image get sent to the central server for deeper analysis. The system includes a ‘Fact Verification’ tool that, for news-related images, cross-references the image content against major news sites including the Times of India using Google search tools and Gemini AI. The training dataset exceeded 49,000 images from celebrity faces to news photographs. The model achieved 82 percent accuracy without compromising user data privacy. Training was completed on a workstation with a 4050-class graphics card in 15 hours of batch processing.

Three papers addressed engineering applications at the intersection of machine learning and physical infrastructure.

Mr. Ashish Shah presented work on intelligent fault detection in DC microgrids. The system under study integrates a photovoltaic power source, battery storage, a grid connection point, and a steady-state load. Faults were simulated to produce realistic electrical signals. The methodology uses discrete wavelet decomposition to extract time-frequency features (entropy, energy, mean, skewness, kurtosis) from the fault signals. These features feed into multiple ML classifiers including Support Vector Machine, Random Forest, Neural Networks, and K-Nearest Neighbour. Random Forest produced the highest accuracy across fault types including open-circuit faults, ground faults, and pole-to-pole faults. Students in the B.Tech in Electronics and Communication Engineering study the signal processing foundations that underlie this kind of fault detection.

Ms. Keyaben Sanket Kumar Patel, research scholar, presented work on AQI prediction for the Dwarka area of Delhi. The dataset spans 2017 to 2024, covering 2,920 days of daily observations. Where conventional AQI predictions use seven major pollutant readings, this work uses 25 weather and atmospheric parameters simultaneously, including wind speed and humidity. Multiple machine learning algorithms were compared: Decision Trees, K-Means clustering, and Support Vector Machines. Random Forest emerged as the strongest performer across all parameters. The evaluator questioned whether multi-class air quality classification (rather than simple binary good/bad classification) was necessary, and the researcher demonstrated that multi-class output provides materially more precise health-relevant information for citizens.

Ms. Ashwini Panase presented work co-authored with Amit Barve and Rajib Chattopadhyay across Parul University’s Department of Computer Science and Engineering, the Faculty of Communication Engineering at the Military College of Telecommunications Engineering, Mhow, and the Indian Institute of Tropical Meteorology, Pune. The case study site was the Mandleshwar gauging station on the Narmada River in Madhya Pradesh, a location with operational significance because its discharge data informs flood warnings for downstream communities including those near the Sardar Sarovar reservoir. Daily discharge data from 2000 to 2018 was used. Ten input features were constructed including lagged discharge values, rolling means and standard deviations, cyclical day-of-year encodings, and binary monsoon-period flags. Four models were compared: ARIMA(5,1,0) as the classical baseline, LSTM, BiLSTM, and GRU. GRU outperformed all others with Nash-Sutcliffe Efficiency of 0.677, RMSE of 540.5 cumecs, and R-squared of 0.677. ARIMA produced a negative NSE of -0.025, indicating that nonlinear monsoon dynamics defeat linear autoregressive models entirely. GRU was recommended as the default recurrent baseline for operational flood warning systems on Indian rivers.

Also Read: Think about digital security in the advancing technology era

Ms. Chaitali Pankajbhai Bhoi presented work on forecasting stock prices in the Indian IT industry using a multimodal framework that combines numerical financial indicators with textual analysis of Management Discussion and Analysis (MD&A) documents from corporate filings. The architecture comprises three branches: textual analysis, market data, and macroeconomic indicators, with a fusion layer integrating the three. Hypothesised relationships included that positive language in MD&A documents correlates with higher subsequent returns, while risk-related vocabulary reflects market volatility. The evaluator noted the persistent challenge that stock markets are influenced by noise and external factors that defeat any single-model approach.

Ms. Munagala DeviSri from the Department of Computer Science Engineering at Malla Reddy University, Hyderabad, presented FormEase Pro as an accessibility tool for web users with dyslexia. The extension combines real-time spelling correction, grammar validation before form submission, field response suggestions from stored user profile data, single-click form filling, font customisation including OpenDyslexic typeface support, and a Reading Ruler feature that follows the mouse cursor to help users maintain visual focus. The architecture pairs a Chrome browser extension with a Flask backend server. Form data is automatically saved every 30 seconds to allow recovery from crashes. The evaluator noted that handling specialised vocabulary (legal, medical) and variable HTML form structures remain ongoing engineering challenges.

Mr. Kayur Rameshbhai Babhaniya presented integration patterns for AI in iOS mobile applications using Core ML and Create ML in Swift. The architecture shifts from cloud-based ML to on-device ML, providing benefits including low latency, improved privacy through local data processing, and offline operation. The two principal application areas were image classification and text-based predictions, both achieving high accuracy in offline operation. Future directions included federated learning and AR technologies. Practical engineering knowledge of this kind is increasingly valuable for Parul University students entering B.Tech in IT and related programmes.

A research team from the Computer Science Department presented a no-code machine learning framework designed to allow non-programmers to build, evaluate, and use ML models on their own data. The motivation is the structural barrier between domain experts (cardiologists, farmers, educators) and the ML capabilities that could help them. The framework’s six-step pipeline automates Data Upload, Preprocessing (missing-value handling, categorical encoding, normalisation), Feature Selection, Model Selection (training Decision Tree, Random Forest, SVM, KNN, and Logistic Regression in parallel), Hyperparameter Optimisation via grid or random search, and Evaluation with full metrics including accuracy, precision, recall, F1-score, and confusion matrix. Future directions include support for deep learning architectures (CNNs, LSTMs, Transformers), real-time deployment via REST API, drag-and-drop workflow builders, and integrated Explainable AI techniques. The evaluator raised the class imbalance question (medical datasets with 95-percent majority class can be gamed by trivial classifiers) as a critical engineering area for the framework.

Across the twenty-two papers, the breadth of application domains is itself the headline. Healthcare from postpartum mental health to brain tumour imaging. Agriculture from leaf disease to multilingual advisory systems. Education from performance prediction to language-specific question answering. Recommender systems for fashion and films. Computer vision from satellite surveillance to image captioning. Civilian safety in armed conflict, secured wearable health monitoring, deepfake detection. Energy infrastructure from microgrids to river forecasting. Accessibility for dyslexic web users. Mobile ML on Apple’s stack. Democratisation of ML for non-programmers.

What unifies the work is not technique but framing. Across all twenty-two papers, the researchers were not building general-purpose models for academic benchmarks alone. They were applying machine learning to specific real-world problems with measurable consequences. Several of the papers carried explicit social grounding: civilian survival in war zones, accessibility for systematically underserved populations, mental health detection, agricultural advisory for small farmers. The shift in Indian academic AI from technique-focused to application-focused work was visible across the session.

The Faculty of Engineering and Technology at Parul University, which hosted PiCET 2026, anchors this kind of work across the B.Tech in Computer Science Engineering, the B.Tech in AI and Machine Learning, the M.Tech in Computer Science Engineering, and the Ph.D. in Engineering and Technology. The new Lakshya 2047 fifteen-laboratory infrastructure provides the computational and software environment that contemporary applied ML research depends on.