Tata Consultancy Services Patente

The present disclosure leverages a graph-based digital twin network and generative artificial intelligence for solving problems of generalized and static risk assessments to create a dynamic risk profile by combining intrinsic farm risks with propagated risks from neighboring farms. This ensures that risk assessments reflect both local and networked influences. The dynamic risk profile is continuously adapted based on real-time data and interconnected farm relationships. Each farm is modeled as a node in the graph-based digital twin network, where edges represent risk propagation pathways influenced by factors such as pest migration, shared resources, and weather dependencies. Further, a real-time, dynamic premium recalibration is implemented for addressing inefficiencies of static premium model, based on the dynamic risk profile, compliance of farm users with risk mitigation advisory recommendations, and unforeseen risk factors. Furthermore, claims validation is performed by integrating remote sensing, IoT sensor data, and crowd-sourced inputs to resolve inaccuracies and fraud.

The embodiments of the present disclosure herein address unresolved problems of occurring hallucination when an LLM generates a response to a query despite not having encountered sufficient or relevant training data to accurately address that specific query of Embodiments herein provide a method and system for detecting and mitigating hallucination in a retrieval augmented generation (RAG). Herein, a pipeline is proposed to make a fine-grained analysis of various components in the RAG pipeline where hallucination can occur and how to mitigate it. By extending the standard RAG pipeline, four major components including a relevance check, an attribution check, an LLM reasoning check, a metric check are introduced to perform the fine-grained hallucination detection and mitigation in RAG-based approach.

The present invention relates to the field of model driven engineering. Creation of the purposive machine-processable models is an intellect-intensive activity requiring business domain as well as modeling expertise, which is in short supply. Hence, embodiments of present disclosure provide a method and system for generating purposive models using a Large Language Model (LLM). The method includes receiving a purposive metamodel and associated pattern trees. Each pattern tree has pattern elements including pattern nodes, with associated pattern properties, connected via pattern edges. The pattern elements may be associated with pattern mapping statements which are executed by a pattern interpretation technique and corresponding objects are created in the purposive model. The pattern mapping statements include prompt generation functions which are executed by the pattern interpretation technique to generate prompt and obtain response from the LLM which is used to populate the purposive model.

Existing approaches used for HIF detection utilize non-algebraic operations which add to computational complexity, consume more time, use complex algorithms and require a very high-sampling frequency (10-20 kHz). Embodiments disclosed herein provide a method and system for high-impedance fault (HIF) detection in a network. In this method, by shifting a domain digital data one cycle time at a time, and processing associated data, a plurality of Harmonics-based Algebraic Indexes (HAI) that include a half-cycle time domain digital data of a one-cycle time domain digital data, and a magnitude strength of a plurality of non-harmonic frequencies of each of the plurality of digital current signals is determined, which captures information on associated magnitude strength information. Further, a time domain aggregated harmonics-based algebraic index (AHAI) indicating presence or absence of High-Impedance Fault (HIF) in a network being monitored is generated, and accordingly the presence of absence of HIF is detected.

The present invention relates to the field of document processing. Existing method of knowledge extraction from documents involve multi-step process to classify bounding boxes into plain text, special symbol, mathematical expression or table and then extract its contents. Identifying proper bounding boxes is also a challenge. Thus, embodiments of present disclosure provide a method and system for extracting contents from documents into predefined format. The method initially divides each page of an input document into multiple bounding boxes. Then, each bounding box is cropped and processed via an OCR reader. Completeness of the bounding box is checked, completeness of equation within the bounding box is checked and text within the bounding box is extracted. Same process is repeated for all bounding boxes in all the pages of the input document. Finally, text extracted from all the pages is consolidated into a predefined format.

A method and system for intelligent robotic controller with task failure detection and handling via a multi-level failure handler is disclosed. A validated task plan received from a task planning framework is analyzed to extract set of skills for execution of the task plan. Each skill comprising atomic skills or compound skills. For the set of skills, set of provisioned skill specific behavior tree (BT) nodes are loaded from a set of skill ledgers from a skill library defining execution of the extracted set of skills. BT train is created for the set of skills based on the set of skill ledgers and the set of provisioned skill specific BT nodes. Failure event is detected and handled by multi-level failure handler and replanning of task is routed to robotic controller or task planning framework based on detected failure mapping to the set of preconfigured failures and a failure handling capability.

This disclosure relates generally to a sequence based neural network method for predicting change in stability of proteins upon single point mutations (SPM). The Stateof-the-art methods are unable to capture change in residue environment using protein sequences information alone. The method of the present disclosure considers novel representation of a wild-type protein sequence and a mutant-type protein sequence by considering residue environment in addition to the residue at the position of mutation. Highly attended residues with respect to the position of mutation form residue environment and are identified by leveraging attention mechanism of the Protein Language Model (PLM) used in this study. Thus, the method captures change in the residue neighbour on mutation without using structural information. The method leads to a more accurate prediction of change in free energy (ΔΔG) at the site of mutation. as compared to existing state of the art methods.

With growing digitization, lifestyle has become sedentary for many citizens. This sedentary behaviour has proven to be dangerous to humans, leading to obesity, cancer, psychological disorders, etc. Hence, organizations and government entities/institutions are promoting an active life cycle. However, it is not easy. Use of mobile applications could be an easy approach to start with, but mobile devices worsen the sedentary behaviour. Gamification is a technique by which intended objectives or tasks performed by users are met through games. Present disclosure provides a system and method that implement sensor boards at different locations in an environment for enabling tasks to be performed by users based on calibrated objects in the environment. Calories burnt by the users are computed based on tasks performed and are compared with the target calories to be burnt. Based on the comparison, the speed of the calibrated objects and an associated difficulty level are adjusted.

The present disclosure provides a system and method for performance monitoring of photovoltaic power generation using a reference power prediction model. In the present disclosure, a plurality of environmental data and a plurality of operational data of a PV power generation unit is received and preprocessed. Further, a clean reference PV power data and a reference irradiance value is obtained in absence of a GTI data in the plurality of the environmental data. However, in presence of the GTI data in the plurality of the environmental data, the clean reference PV power data is predicted using a power prediction model. Furthermore, a uniform value of a performance index for the PV generation unit is computed using a performance ratio for a predefined duration using an optimization framework. Performance monitoring of the PV power generation unit is performed by identifying a rising and a dropping pattern of the performance index.

The disclosure generally relates to methods and systems for real-time monitoring and optimization of a packed bed column. Conventional techniques for the packed bed simulation in the literature are very limited which employ fundamental physics-based models such as Ergun and Darcy models in conjunction with fixed feed inlet velocity boundary condition, resulting in erroneous predictions. The present disclosure constructs a digital twin of the physical packed bed column by predicting state parameters in a spatio-temporal domain inside the packed bed column, using first principle-based model, a data-driven model, a knowledge driven model, or a combination thereof. The one or more state variables and state parameters are then processed to determine and optimize the one or more key performing indicators (KPIs) of a process and each of one or more process steps of the process. The obtained digital twin is then used for monitoring the physical packed bed column in real-time.

This disclosure relates generally to a method of selecting a machine learning (ML) model for regional risk assessment of susceptible region. State-of-the-art methods provide various ML models for risk assessment that provides hazard susceptibility maps. However, the selection of an optimal ML model for the susceptible region remains a challenge. The method utilizes an analysis of a plurality of factors (qualitative and quantitative) that are influenced due to change in ecological environment of the suspectable region prior to an onset of the hazard. The method further involves generating an individual scores based on assessment of the plurality of factors, availability of labelled samples, and resolution level of images of the suspectable region. Each individual score is combined to obtain a selection score. The selection score is scanned to a pre-classified ML models to select an optimal ML model suitable for assessment of the expected hazard based on the selection score.

A method and system for computation of molecular properties via quantum autoencoders is disclosed. The disclosure provides quantum machine learning-based method to compress molecular Hamiltonians and efficiently relevant property of a molecule given at a geometry, thereby significantly reducing resource requirements. The method disclosed leverages a quantum autoencoder which is trained using ground states corresponding to molecular geometries of the molecule and a classical optimizer. An n-quantum bit (qubit) transformed property operator of an n- qubit property operator is determined using the trained quantum autoencoder. Further, n-qubit transformed property operator is compressed by performing a series of Pauli term processing to obtain a m-qubit compressed property operator. Further, a relevant molecular property at the geometry is determined using a variational quantum eigen solver (VQE) or using compressed ground state. The present disclosure is used to determine any property of the molecule at a given geometry using a reduced resource requirement.

This disclosure relates generally to a method and system for generating spatially distributed response to query by intelligent selection of time-series remote sensing data. The current available methods for processing and analyzing geospatial data have limitations such as manual satellite data selection, limited natural language query support and the like. The method disclosed receives multi-contextual user query, intelligently selects the relevant time-series satellite data and climate information from repository using intents and entities generated utilizing pre-defined knowledge graph. The method intelligently selects time-series satellite data by creating and utilizing relationship strength-based node ranking graph. This time-series satellite data is combined with the climate data selected from climate repository to generate the spatially distributed response for the multi-contextual user query using machine learning model and large language model. The method enables stakeholders to access actionable insights without needing expertise in satellite data processing and using it for informed agricultural decision-making.

This disclosure relates generally to method and system to determine sustainability risks in multi-tier supplier chains for EUDR compliance. Multi-tier supply chain management suffer from insufficient transparency, challenges in oversight, and often, a deficit of trust. The method receives a purchase order from a user to validate source supply of raw materials procured in accordance with EUDR compliant. In response each supplier provides a list of documents for the purchase order to assess EUDR compliance. Further, spatiotemporal data is obtained from the purchase order to determine a plurality of supplier chain risk assessment factors using a multi-tier supply chain artificial intelligence (AI) model. Finally, a compliance determination model determines at least one sustainability risks for EUDR compliance to differentiate suppliers risk based on a plurality of risk assessment factors to classify into at least one of compliant, non-compliant or deemed non-compliant.

The disclosure relates generally to methods and systems for cross-domain based change detection of region due to an event. Conventional techniques on cross-domain change detection (CDCD) rely on transformation-based approaches involving two tasks: image translation from one modal to another modal (SAR to optical) and then performing the change detection (CD) in one of the translated modals (for example, in SAR or optical). The methods and system of the present disclosure propose a deep-learning (DL) architecture called ReFUjetNet leverages generic embeddings from the extensively pre-trained domain adaptive foundation model alongside locally trained embeddings. The ReFUjetNet efficiently fuses these embeddings and employs a Kolmogorov-Arnold Network (KAN)-based convolutional neural network (CNN) classifier to generate a binary change matrix, which is then joined with another binary change matrix derived from the Gabor jet-based dissimilarity checker, resulting in the final binary change map.

This disclosure relates generally to method and system for quality assessment of physical grade tea blend using a dual segmentation technique. Tea blending quality assessments rely on manual inspection methods, which involves subjective visual evaluation of particle appearance, size, and composition. To address the challenges and maintain tea blend consistencies, the method employs a dual segmentation technique to extract spatial features and spectral features from a made-tea sample thereby assessing uniform distribution of tea particles across various grades in a physical grade tea factory. Further, bounding boxes are marked on spatial segmentation masks and spectral segmentation masks ensuring the tea blend is thoroughly evaluated without overlaps or missing key regions. Then, a blending quality score is computed which evaluates accuracy and consistencies in tea blends. The method further generates a blending quality assessment report recommending the tea factory indicating personalized improvements to adjust blending process based on the blending consistency.

Using human pose detection technology, a system and method disclosed herein tracks poses performed by a subject and compares them with expected poses in an educational game. As the subject performs these actions, they are presented with math-related tasks, improving both engagement and learning through active participation. The method generates random numbers and creates combinations, challenging subjects to make correct poses based on the onscreen instructions. The response of the subject is evaluated using a unique scoring model and dynamically toughness of the questions asked is varied based on the score. The response scoring model is based on nonlinear scoring where score is provided for speed, complexity, incorrectness and confidence of subject while responding to questions. Existing solutions that analyze response are focused on mere classification of confidence levels and are generic in classification approaches using ML models, which are computation heavy and limited to mere class based classification.

For developing recipes for new products in the lyophilization system, extensive resource-consuming experiments need to be carried out. The parameters obtained through such experiments are often sub-optimal. Present disclosure provides a system and a method for predictive control of a lyophilization process by capturing the effect that freezing has on the product characteristics and consequently on the drying stages. The system utilizes real-time input from a lyophilization system, and the state estimation models to closely and predictively control the desired attributes of the product and determine optimal operating parameters for a desired product quality based on a comparison of the current state of the lyophilization system with a reference state. Image of the product is captured to identify defects in the vials, trace back to operating conditions and generate recommendations for defect mitigation.

Unlike natural forests, urban forestry presents a multifaceted approach for addressing climate change. However, these are extremely vulnerable to the impacts of climate change and severely hampered if forests do not follow sustainable forest management or are overused. Further, measuring and monitoring climate smartness of urban forest have certain challenges. Present disclosure provides systems and methods for real-time monitoring of climate smartness of an urban forest region. The climate smartness of the urban forest region is assessed using satellite and aerial images to monitor urban forest health indicators, field data on tree health, soil quality and biodiversity. Biomonitoring is done using indicator species to assesses ecological health of the urban forest region. Various indicators are computed to assesses the smartness of the urban forest using index that represents as an overall score to measure and monitor climate smartness of an urban forest in meeting the requirement of sustainable forest management.

Epitopes identification is a very important problem and has implications in designing vaccines for emerging viruses, bacteria and fight antimicrobial resistance. Traditional methods of epitope mapping are costly, and time consuming. Present disclosure provides a system and a method that processes structural features along with the features from protein language models and apply deep concatenation techniques to combine the features for obtaining a concatenated feature set. A weight for each feature amongst the concatenated features set and a bias for the concatenated features set are learnt to obtain an optimal combination of features set based on which amino acid residues are predicted as one of an epitope or a non-epitope. Further, epitopes are prioritized using a plurality of attention scores capturing a T-B association by using a protein language model, to obtain a set of immunogenic B-cell epitopes.

The disclosure relates generally to methods and systems for recommending adaptive crop protocol without adversely impacting yield and emissions. Conventional techniques in crop protocol recommendation may not always take goals of the farmers into consideration and hence may not provide balance between the yield, water usage and the emissions. The present disclosure solves the technical problems in the art by balancing the agriculture yield with environmental sustainability, ensuring that productivity does not come at the cost of environment while the crop protocol is recommended. The present disclosure recommends the adaptive crop protocol of the crop to be cultivated by the user, based on the plurality of transformed user goals, using a trained ensemble model and an agriculture operation sequencing technique. Further, a Yield At Cost of Environment (YACE) index is determined which helps in assessing an environment sustainability of the recommended adaptive crop protocol.

A method and system for real time automated grading of tea shoots for quality and resource management is disclosed. The method disclosed implements a multi-layer data fusion approach integrating RGB and spectral data of tea leaves for simultaneous morphological and chemical profiling. The method provides real-time detection of critical anomalies and an anomaly scoring mechanism. The method links real-time chemical profiling data with flavor prediction models and adjusts processing parameters based on target flavor profile. The method provides a dynamic worker compensation mechanism which includes real-time compensation to leaf quality metrics, incorporating post-segregation anomaly handling and rewarding performance based on output quality. The method allocates rejected leaves to optimized repurpose methods, minimizing waste and promoting sustainability. Thus, the method disclosed enables real-time tea leaf classification, flavor profile prediction, manufacturing optimization, dynamic worker compensation, waste management, and comprehensive reporting in tea manufacturing industry.

Existing techniques for ontology generation and extraction associated rules typically involving custom pattern-matching algorithms are complex and require significant human intervention to design and implement. Embodiments herein provide a method and system for generation of the ontology and extracting the associated rules using meta-models. The method comprises user defining a domain using the meta-model in a JavaScript Object Notation (JSON) format for tailored and precise generation of the ontology. A plurality of prompts is generated using the meta-model, a configuration file, and a domain specific corpus, and processed using a large language model (LLM) to generate the ontology. Further the generated ontology along with the domain specific corpus is used to extract associated rules effectively.

Conventional techniques for multi-modal image segmentation are not computationally effective for image analysis. Embodiments herein provide a method and system for an organ prioritized Y-Net multi-modal image segmentation model using a joint modality encoder and decoders. The method trains a genetic Y-Net multi-modal image segmentation model comprising the joint modality encoder, a bottleneck, and a plurality of modality-specific decoders, with each of a plurality of the multi-modal medical images, for a plurality of predefined hyperparameters, to generate a trained Y-Net multi-modal image segmentation model. Further a multi-objective organ prioritization reward function with an organ prioritization is formulated for the plurality of predefined hyperparameters of the trained Y-Net multi-modal image segmentation model. Further the plurality of parameters of the trained Y-Net multi-modal image segmentation model is optimized using the multi-objective organ prioritization reward function, to generate the organ prioritized Y-Net multi-modal image segmentation model.

This disclosure relates generally to method and system for made-tea quality assessment in manufacturing process. Conventional tea quality assessment methods making it challenging for tea manufacturers to maintain consistent quality standards and meet evolving market demands. The present disclosure collects a physical grade of made-tea sample from the tea factory. Spectral camera captures spectral image of the made-tea sample for preprocessing to obtain improved contrasts. Further, from a plurality of features are extracted from each candidate region to detect a plurality of anomalies, and simultaneously a negative score is assigned based on detected discrepancies. Then, a quality level score is computed to determine a final grade value. The final grade value is assessed based on a quality level score (QLS) and a valuation grade score (VLS). Finally, the method generates a quality assessment report indicating personalized improvements for the tea factory.

While knowledge distillation has shown promise in general NLP contexts, its application to domain-specific text remains largely underexplored. Existing models often struggle to generalize complex domain-specific vocabulary, specialized terminology, and unique sentence structures, resulting in degraded performance when applied to domain-specific tasks. Thus, embodiments of present disclosure provide a method and system for domain-specific text processing with keyword knowledge distillation. The disclosed method overcomes the challenges in conventional methods by utilizing a dynamic, weighted ensemble approach for keyword extraction, which adaptively selects domain-specific keywords extracted using multiple keyword extraction techniques. The method then applies knowledge distillation, where a student model is trained to predict the masked domain-specific keywords using context from surrounding tokens. By computing three loss functions, including distillation loss, keyword masking loss, and cosine embedding loss, the student model is optimized for performance while retaining critical domain knowledge.

A method and system for contextual search on a database based on dynamic reward function using relevancy count is disclosed. Generative artificial intelligence has challenges in handling complex natural language queries for database transactions that require contextual rather than direct column mappings. The method disclosed extracts contextually relevant data from relational databases using a combination of Retrieval-Augmented Generation (RAG) and a contextual Monte Carlo Tree Search (CMCTS) with a dynamic reward function. The method disclosed uses the dynamic reward function during the simulation step of the CMCTS on a Monte Carlo Tree by calculating a contextually relevant score of a current node in the Monte Carlo Tree. The method also uses a modified upper confidence bound during selection step of the CMCTS. The method disclosed solves the real-world problem of searching for companies relevant to a given topic or capability.

Conventionally, applications using machine learning models apply same models throughout the working of the application but there is a need to select and switch to appropriate models based on the input data for providing better results. Hence, embodiments of present disclosure provide a method and system for dynamic selection of machine learning models. The method receives input data, extracts features from it and derives a set of parameters from the features. Then, a model fit vector is generated as a weighted combination of the set of parameters. The weights are determined using an entropy-SHAP hybrid technique and further fine-tuned based on a fairness metric. Next, machine learning models are trained using the parameters having weights greater than a threshold value and a model score is calculated for each of the models. Best model is selected based on the model score.

Industrial structures are prone to external conditions resulting in leakage, faults etc. While easy-to install structure monitoring methods exist, sensor maintenance and in-situ inspection becomes arduous, if the structure is at an inaccessible location. For example, localization of leak event and the ability to differentiate leak and non-leak becomes an even more daunting challenge. Method and system disclosed herein facilitate wideband beam forming based target localization. Further to the wideband forming, the system identifies and processes a single channel audio frame from an identified microphone that is nearest to an angle of interest to generate a spatially filtered single channel audio frame. Further, the spatially filtered single channel audio frame is fused with a set of single frequency acoustic maps. Further, a target localization of an event of interest in the angle of interest is performed by fusing the spatial-spectral filtered single channel audio frame with the composite acoustic map.

The embodiments of the present disclosure herein address unresolved problems in smart mirrors available at different points of retails shops. Existing mirrors never have a try to really understand the preference and interest of a user. Also, the existing mirrors are not up to mark providing a voice based chat engine which can be carried in the handbags easily. Embodiments herein provide a method and system for a smart makeup mirror by combining a large language model (LLM) with hardware components. By combining the LLM with hardware components, this device creates an interactive beauty assistant that provides personalized advice capabilities. The system is configured to create an interactive questionnaire which helps someone to identify themselves with their taste and interest in make up in a personalized way of interactive questionnaire process and an advice engine explicitly for makeup.

The disclosure relates generally to methods and systems for determining optimized cloud architecture using a modelling-based approach. Most of the conventional techniques for cloud architecture design pose limited capabilities. The methods and systems of the present disclosure introduce an end-to-end framework consisting of a workload-aware, model-based approach for designing optimal cloud deployment architectures, and for recommending cloud architecture designs tailored to deploying applications in the cloud. The service-specific parameters are fed into underlying cost models and service specific performance models, which then generate individual service and end-to-end deployment cost and performance estimates. Once the basic architecture is built, the user can perform what-if and if- what analysis. The impact on performance (performance estimation) and cost (cost estimation) by replacing services or changing the configurations can be analyzed using the models in the user dashboard to design the optimized cloud architecture.

A method and system for Al based recipe generation by integrating ingredient pairing score determined from flavor spaces is disclosed. Current available recipe generation systems do not focus on creating recipes optimized for the specific nutritional needs associated with various health conditions. The method disclosed incorporates nutritional and flavor profiles for pairing suitable ingredients and generating recipes. The method identifies a primary ingredient based on conditional nutrient-based filtering technique. Based on the selected primary ingredient the method employs flavor compound-based embeddings to assess flavor compatibility between this primary ingredient and other potential ingredients, identifying suitable secondary ingredients for the recipe. The secondary ingredients are determined by determining ingredient pairing score using directional similarity measure, orientation factor and spatial proximity measure from molecular embedding matrices represented as flavor spaces of the primary ingredient and the secondary ingredients. The method generates personalized recipes used to manage various lifestyle related diseases.

Exiting systems fail to provide an automatic way of assessing the application with respect to crypto agility. The present disclosure performs the following steps iteratively, till the length of a call hierarchy corresponding to one or more interfaces of one or more crypto Application Programming Interfaces (APIs). A crypto inventory report is generated using a first set of inventory rules. The one or more clusters of one or more crypto APIs are generated based on a membership of one or more crypto API parameters of a class determined by a slicer. The one or more interfaces are extracted from the one or more clusters of the one or more crypto APIs. The extracted one or more interfaces are converted into a second set of inventory rules and a crypto inventory report is generated. An agility score of the application is then calculated accordingly.

Current emotional detection technologies often struggle to consistently identify emotional state shifts in real-time, particularly in noisy environments where signals fluctuate due to minor external changes, and also fail to provide a robust mechanism that can account for and filter out these transient signals. Method and system disclosed herein perform real-time emotional state shift detection of a user and adaptive response generation. The system captures and processes a multi-modal input data in real time, during which weight of each multi-modal input data is determined based on a sensitivity factor that is determined by performing dynamic sensitivity adjustment for each multi-modal input data. An adaptive question is generated for the facilitator and the response is collected. The user is made to play a generative game until the user is identified to be in a pre-defined state. Further, a customized interview structure is generated for the user.

The embodiments of the present disclosure herein address unresolved problems of existing idea generation approaches depart from the realistic setting, where the key challenge lies in prioritizing retrieval of the literature most conducive to generating novel research ideas, provided only the research problem and its motivation as the query. Embodiments herein provide a method and system for methodology inspiration retrieval for scientific research problems. The system leverages a rich citation network within a research domain to derive a directed Methodology Adjacency Graph (MAG). The directed MAG captures methodological lineage, with the corresponding citation contexts from the citing papers. Further, a dataset is curated to fine-tune retrievers and evaluate MIR task by extending multi-cite dataset. To fine-tune a more effective retrieval model for MIR, a triplet loss is formulated by synthesizing samples from the directed MAG. Further, a reasoning-intensive process essential for MIR is simulated by extending a predefined LLM-based re-ranker.

The ability to generate contextually relevant multimodal answers based on proprietary knowledge base is critical for enhancing user experience in interactive systems. Conventional approaches fails in generating precise, multimodal responses. The present disclosure provides a system and method for persona-specific, recursive clarification in Large Language Model (LLM)-driven query response generation. The system enhances the LLM's interaction process by generating tailored clarification questions based on a Multi-layer hierarchical relational graph (MHRG) that profiles user intent and maps domain-specific knowledge relationships. By iteratively refining and clarifying the user's question, the system generates a user's intent profile that optimizes the retrieval and response process, enabling the LLM to better address complex or ambiguous queries. This framework ensures that each interaction dynamically adapts to the user's intent and persona, improving interpretive accuracy and engagement.

Existing solutions often rely on linearization or heuristic methods, which fail to capture the nonlinear interdependencies inherent in promotion planning. A method and system for optimizing retail promotions with Mixed-Integer Non-Linear Programming (MINLP) and cross-elasticity effects is disclosed. The method uses two-step heuristics and Mixed-Integer Nonlinear Programming (MINLP) for promotion planning. The promotion optimization problem is to identify the optimal selection of promotional items and to obtain the optimal discounts for the set of the selected products or items and bundles, so that the weighted combination of total revenue and the total margin at a category level is maximized for the promotion period. Further, it achieves the given total revenue and total margin target numbers, while satisfying the constraints including minimum and maximum discounts, total budget, bundles conflicts and the business rules. Retailers require advanced demand forecasting tools that integrate historical sales data, price elasticity, budget constraints, and operational considerations.

Multi-sensor fusion combines data from different sensors to improve accuracy and reliability. Existing methods for multi-sensor fusion utilize Deep Neural Networks which need large, labeled datasets and heavy computational resources. Hence, present disclosure provide a method and system for multi-sensor data fusion using contrastive learning based deep convolutional transform learning for limited data. The method collects data from multiple sensors which is processed by Deep Convolutional Transform Learning (DCTL) encoders, each specific to a sensor, to extract sensor-specific features. Further, a DCTL loss is computed for each DCTL encoder, and a contrastive loss is computed between all feature pairs generated by the DCTL encoders. The method jointly optimizes parameters of the sensor-specific DCTL encoders and contrastive loss function to extract fused features effectively, enabling robust multi-sensor fusion with limited data.

Data generated by LLMs require to be validated to ensure accuracy. To address this issue, the generated data had to be manually validated, which was time-consuming and impractical as the volume of generated data could be huge. Though there are some automated approaches to provide a good starting point for assessing the quality of the generated data, these approaches do not correct the data or explain reasons for the errors in the data. Method and system disclosed herein facilitate fine-tuning of a validator, using a formal clause based approach. The validator, once fine-tunes, is then used for identifying a generator that generates accurate data values, which in turn maybe used for different applications.

Existing design methods rely on manual effort, trial-and-error approaches, and simplified models, which can lead to suboptimal designs and reduced efficiency. The present disclosure obtains a first set of responses for a first set of questions from a user and generates a second set of questions. A second set of responses is obtained and one or more questions specific to a problem formulation are generated using one or more autonomous programs. A first set of data corresponding to one or more questions is retrieved from a knowledge base and one or more gaps is identified. A second set of data is retrieved from one or more resources and knowledge base is updated. One or more process models and corresponding one or more optimization models are retrieved, and a relevant process model is selected. The most appropriate one or more optimization models are selected and, and feedback is obtained from user.