Issuu

Style Wise: An AI-Powered E-Commerce SaaS Platform with Hybrid Recommendation Engine and Conversational Voice Assistant

Joothiswaran1 , Alavudeen2 , Deepan3, Karthik4

Bachelor of Technology in Artificial Intelligence and Data Science Suguna College of Engineering, Coimbatore, Tamil Nadu, India.

Abstract - The rapid growth of online fashion retail has intensified the demand for intelligent product discovery systems that can understand individual user preferences and delivercontextuallyrelevantsuggestions.Thispaperpresents StyleWise, an AI-powered e-commerce Software-as-a-Service (SaaS) platform that integrates a hybrid recommendation engine with a conversational voice assistant to deliver a personalized shopping experience. The recommendation engineemploysascalabletwo-stagearchitectureconsistingof candidategenerationfollowedbymulti-factorranking.During candidategeneration,fourindependentstrategies categorybased retrieval, content similarity matching, localized collaborativefiltering,andtime-decayedtrendinganalysis reduce the product space from thousands of items to approximately 200 candidates. A hybrid ranking engine then scores each candidate using a weighted combination of five signals: intent match (0.35), preference alignment (0.25), collaborative boost (0.20), popularity (0.10), and recency (0.10). The platform further incorporates a GPT-4 powered voice chatbot that enables natural language product search, cart management, and order tracking through speech recognitionandtext-to-speechsynthesis.Thesystemisbuilton amicroservicesarchitectureusingFastAPI,Next.js,MongoDB, and Redis, containerized with Docker for horizontal scalability. Experimental evaluation shows that the recommendation engine achieves a response latency of approximately45millisecondswithamemoryfootprintunder 150megabytes,supportingover500requestsper second.The platform gracefully handles the cold-start problem through onboarding preference collection and trending-product fallback strategies.

Key Words: AI Recommendation Engine, E-Commerce Platform, Hybrid Scoring Algorithm, Collaborative Filtering, Voice Chatbot, Microservices Architecture, Cold-Start Problem, Personalization

1. INTRODUCTION

The global e-commerce market has witnessed unprecedentedgrowth,withonlinefashionretailbecoming oneofthemostcompetitivesegments.Accordingtorecent industryreports,consumersarefrequentlyoverwhelmedby thesheervolumeofavailableproducts,leadingtodecision fatigue and abandoned shopping sessions. Traditional ecommerce platforms rely on rule-based filtering and keyword search, which fail to capture the nuanced preferences of individual shoppers. This gap between

productavailabilityandproductdiscoverabilityhascreated apressingneedforintelligentrecommendationsystemsthat can understand user intent and deliver contextually appropriatesuggestions.

Existing recommendation approaches, while effective in specificdomains,presentseverallimitationswhenappliedto fashion e-commerce. Content-based filtering systems strugglewiththesubjectivenatureoffashionpreferences. Collaborativefilteringmethodsfacescalabilitychallengesas they typically require loading the entire user-item interaction matrix into memory. Furthermore, most commercialplatformstreattherecommendationengineand user interaction layer as separate concerns, missing the opportunity to create a unified, conversational shopping experience.

ThispaperpresentsStyleWise,anend-to-endAI-poweredecommerce SaaS platform that addresses these challenges through three key innovations: (1) a two-stage recommendation architecture that achieves sub-50 millisecondresponsetimeswithoutloadingthefullproduct dataset into memory, (2) a hybrid five-factor scoring algorithm that combines short-term browsing intent with long-term preference patterns, and (3) a GPT-4 powered conversational voice assistant that enables hands-free productdiscoveryandshoppingactions.

Theremainderofthispaperisorganizedasfollows.Section 2 reviews related work in recommendation systems and conversational commerce. Section 3 details the system architecture and design methodology. Section 4 describes theimplementationoftherecommendationengine.Section5 covers the voice chatbot subsystem. Section 6 presents experimentalresultsandperformanceevaluation.Section7 concludesthepaperwithadiscussionofcontributionsand futuredirections.

1.1 Problem Statement

Modern e-commerce platforms face a fundamental challenge:asproductcatalogsgrowtothousandsormillions ofitems,usersstruggletodiscoverproductsthatalignwith their personal style, budget constraints, and current shoppingintent.Traditionalrecommendationsystemssuffer fromthreecoreproblems.First,theyrequireprohibitively expensivematrixfactorizationoperationsthatdonotscale beyond moderate catalog sizes. Second, they fail to

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 02 | Feb 2026 www.irjet.net p-ISSN: 2395-0072

distinguishbetweenauser'simmediatebrowsingintentand their long-term preferences. Third, they provide no mechanism for conversational product discovery, forcing userstointeractthroughrigidsearchinterfaces.

1.2 Objectives

Theprimaryobjectivesofthisresearchareasfollows:

 To design and implement a scalable two-stage recommendationenginethatoperatesonper-user indexedqueriesratherthanfull-datasetloading.

 To develop a hybrid ranking algorithm that integratesfivedistinctrelevancesignalstoproduce explainableproductsuggestions.

 Tobuildaconversationalvoiceassistantcapableof understandingnaturallanguageshoppingqueries andexecutinge-commerceactions.

 To architect a multi-tenant SaaS platform using containerized microservices for horizontal scalability.

 Toevaluatethesystem'sperformanceintermsof latency, throughput, memory efficiency, and coldstarthandling.

 1.3Scope

Thescope ofthiswork encompassesthecompletedesign, implementation, and evaluation of an AI-powered ecommerceplatformtargetingthefashionretaildomain.The platform supports user registration and authentication, product catalog management, personalized recommendations,voice-basedshoppinginteractions,cart andordermanagement,andadministrativeanalytics.

2. LITERATURE REVIEW

Recommendationsystemshavebeenextensivelystudiedin theinformationretrievalandmachinelearningcommunities. Thissectionreviewsthemajorapproachesandpositionsthe contributions of this work within the existing body of knowledge.

2.1 Content-Based Filtering

Content-basedfilteringrecommendsitemssimilartothosea userhaspreviouslyinteractedwith,basedonitemattributes [1].Infashione-commerce,itemattributesincludecategory, colour,material,pricerange,andstyledescriptors.Lopset al. [2] demonstrated that content-based approaches work wellwhenrichitemmetadataisavailablebutstrugglewith the serendipity problem, where users receive recommendations that are too similar to their existing preferences. Style Wise addresses this limitation by combiningcontentsimilaritywithcollaborativeandtrending signals.

2.2 Collaborative Filtering

Collaborative filtering identifies patterns across user behavior to recommend items that similar users have

enjoyed[3].Traditionalmatrixfactorizationtechniquessuch asSingularValueDecomposition(SVD)requireconstructing thefulluser-iteminteractionmatrix,whichgrowsasO(n× m)wherenisthenumberofusersandmisthenumberof products. Koren et al. [4] proposed several optimizations, but the fundamental memory constraint remains. The localizedcollaborativefilteringapproachinStyleWiseavoids thisbottleneck byqueryingonlytheinteractionhistoryof users who share product purchases with the target user, achieving O(k) complexity where k is the size of the individualuser'sinteractionhistory.

2.3 Hybrid Recommendation Systems

Hybrid systems combine multiple recommendation strategies to overcome the weaknesses of individual approaches[5].Burke[6]categorizedhybridmethodsinto weighted,switching,mixed,andcascadedesigns.Industrial systems such as YouTube's recommendation engine [7] employatwo-stagearchitecturewithcandidategeneration followed by ranking, which has become the standard for large-scalesystems.StyleWiseadoptsthistwo-stagedesign whileintroducingafive-factorscoringformulaspecifically tailoredforfashione-commerce.

2.4 Cold-Start Problem

Thecold-startproblemoccurswheninsufficientinteraction data is available for new users or new products [8]. Approaches to mitigate cold start include onboarding questionnaires,popularity-basedfallbacks,andknowledgebased recommendations. StyleWise employs a dynamic weight-shiftingstrategywherethetrendingproductsignal weightincreasesfrom0.15to0.55foruserswithfewerthan threerecordedinteractions,ensuringthatnewusersreceive meaningfulrecommendationsfromthefirstsession.

2.5 Conversational Commerce

Conversational commerce leverages chatbots and voice assistants to facilitate shopping interactions [9]. Recent advances in large language models (LLMs) such as GPT-4 have enabled more natural and context-aware conversationalagents.However,integratingLLM-powered chatbots with live e-commerce backends for real-time product search, cart operations, and order management remainsanactiveareaofresearch.StyleWisecontributesto thisspacebyimplementingafull-stackvoicechatbotwith intentdetection,entityextraction,dialogstatemanagement, andactionexecutioncapabilities.

3. SYSTEM ARCHITECTURE

ThissectiondescribestheoverallarchitectureoftheStyle Wise platform, covering the client layer, API gateway, microservices,AIservices,anddatalayer.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 02 | Feb 2026 www.irjet.net p-ISSN: 2395-0072

3.1 Architectural Overview

StyleWisefollowsamicroservices-basedarchitecturewith clearseparationofconcerns.Thesystemisdividedintothe followinglayers:

Table 1: Platformtechnologystackbyarchitecturallayer

Layer Components Technology

ClientLayer Web Frontend, AdminDashboard Next.js16,React 19,TypeScript

Authentication User Identity Management Firebase Authentication

APIGateway Routing, JWT Validation, Rate Limiting FastAPI,Redis

Microservices Product, Order, Cart,UserServices FastAPI(Python 3.11+)

AIServices Recommendation Engine, Voice Chatbot FastAPI, scikitlearn,GPT-4

DataLayer Primary Database, Cache,Search MongoDB 7.0, Redis

Infrastructure Containerization, Orchestration Docker, Docker Compose

3.2 Frontend Architecture

Theclient-facingapplicationisdevelopedusingNext.js16 with the App Router paradigm, leveraging React 19 and TypeScript for type-safe component development. The frontend employs a component library built on Radix UI primitives with Tailwind CSS for styling. Key frontend modulesinclude:

 Product Discovery Interface:Displays AIrecommended products with explanation cards showingwhyeachitemwassuggested.

 VoiceChatbotWidget:Afloatingchatinterfacewith real-timespeechrecognitionusingtheWebSpeech APIandanimatedwaveformvisualization.

 Onboarding Flow:Collects user style preferences (Minimalist, Classic, Trendy), budget range, occasion preferences, and color choices during initial registration to address the cold-start problem.

 AdminDashboard:Providesanalyticsvisualizations, user management, and product catalog administrationusingRechartsfordatavisualization.

3.3 API Gateway

TheAPIgatewayservesasthesingleentrypointforallclient requests.ItperformsFirebaseJWTtokenverification,issues platform-specificJWTtokenswithconfigurableexpiration, resolves tenant context for multi-tenant isolation, applies role-based rate limiting (30 requests per minute for anonymoususers,100forcustomers,300forsellers,1000 foradministrators),androutesrequeststotheappropriate downstreammicroservice.

3.4 Microservices Layer

Theplatformcomprisesfivecoremicroservices,eachwitha singleresponsibility:

 ProductService(Port8001):Managestheproduct catalog with full CRUD operations, category management, search functionality, and inventory tracking.

 Order Service (Port 8002):Handles the complete order lifecycle including checkout, payment processing,fulfillmenttracking,andorderhistory.

 AIRecommendationService(Port8003):Hoststhe two-stage recommendation engine described in Section4.

 Chatbot Service (Port 8004):Hosts the voice chatbotwithGPT-4integrationdescribedinSection 5.

 Admin Service (Port 8005):Provides tenant management, analytics aggregation, and platform configuration.

Inter-service communication uses synchronous REST for client-facing operations and Redis Pub/Sub for asynchronouseventpropagation(orderplacedevents,user behaviourevents,inventoryupdates).

3.5 Data Layer

MongoDB7.0servesastheprimarydatabase,storinguser profiles,productcatalogs,orders,carts,andAIinteraction logs.MongoDBwasselectedforitsflexibledocumentmodel, which accommodates the varying attribute schemas of fashionproductsacrosscategories.Redisprovidessession management, caching of recommendation results, rate limiting counters, and event bus functionality through Pub/Sub. SQLite with indexed queries serves as the local datalayerfortheAIrecommendationengine,enablingO(log n)lookupsonproductandinteractiontables.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 02 | Feb 2026 www.irjet.net p-ISSN: 2395-0072

4. AI RECOMMENDATION ENGINE

Therecommendationengineisthecoreintelligencelayerof StyleWise.Itemploysatwo-stagearchitectureinspiredby industrialrecommendationsystems,adaptedforfashionecommercewithdomain-specificscoringsignals.

4.1 Stage 1: User Modeling

User modeling extracts a comprehensive profile from two temporalperspectives:

Short-Term Intent Extraction:Captures the user's current browsing session behavior by analyzing the last 10 interactions. A recency decay function assigns higher weightstomorerecentlyviewedproducts:

whereiisthepositionintheinteractionhistory(0=most recent) and λ = 0.3 is the decay rate. This produces a weighted product list, from which recent categories, keywords,andproductsimilaritysignalsareextracted. Long-Term Preference Extraction:Aggregates historical behaviorovera90-daywindowtoidentifystablepreference patterns. The system computes category affinity scores (normalizedto[0,1]),preferredpricerange,averageorder value, purchase frequency, and a ranked list of preferred productsbasedonrepeatpurchasebehavior.

4.2 Stage 2: Candidate Generation

Candidate generation reduces the product space from thousandsofitemstoapproximately200candidatesusing fourindependentretrievalstrategies:

Source1–Category-BasedRetrieval(25%weight):Fetches topproductsfromtheuser'srecentlybrowsedcategories. Categoriesarerankedbyrecency,andproductswithineach categoryaresortedbyrelevancescore.Thissourcetargets userswithclearcategory-levelpreferences.

Source 2 – Content Similarity (25% weight):Identifies products similar to the user's most recently viewed items using keyword-based matching. Product names and descriptionsaretokenized,andaterm-frequencysimilarity measureretrievestheclosestmatches.Thissourcecaptures fine-grainedstylepreferenceswithincategories.

Source 3 – Localized Collaborative Filtering (25% weight):Insteadofconstructingthefulluser-itemmatrix,the systemidentifiesuserswhopurchasedthesameproductsas the target user (limited to the top 20 similar users per product),thenretrievesproductspurchasedbythosesimilar users. This approach achieves O(k) complexity per user ratherthanO(n×m)forthefullmatrix.

Source 4 – Trending Products (25% weight):Computes a time-decayedtrendingscoreforproductsbasedonrecent purchasevelocity:

whereP7disthepurchasecountinthelast7days,P30dis the30-daycount,andVisthevelocitycalculatedas(P7d/7) − (P30d/ 30). This source is essential for addressing the cold-start problem and for introducing serendipitous discoveries.

4.3 Stage 3: Hybrid Ranking

Therankingenginescoreseachcandidateusingaweighted linearcombinationoffivenormalizedsignals:

Table 2:Rankingsignalweightsanddescriptions

Signal

IntentMatch (Sintent) 0.35

Preference Match(Spref) 0.25

Collaborative Boost(Scf) 0.20

Alignment with current browsing session categories, keywords, and viewedproducts

Alignment with long-term categoryaffinityandprice rangepreferences

Scorederivedfromsimilaruserpurchaseoverlap

Popularity (Spop) 0.10 Time-decayed trending score

Recency Boost (Srec) 0.10

Exponential decay based on time since last interaction

Therecencyboostforeachcandidateiscalculatedas:

wherehisthehourssincetheproductwaslastinteracted withandh1/2=24hoursisthehalf-lifeparameter.

Research

Volume: 13 Issue: 02 | Feb 2026 www.irjet.net

4.4 Cold-Start Handling

Foruserswithfewerthanthreerecordedinteractions,the system dynamically redistributes the candidate source weights:

Table 3: Dynamicweightshiftingforcold-startusers

Source

Additionally,theonboardingflowcollectsexplicitpreference signals(style,budget,occasion,colors)thatareusedtofilter andboostcandidatesevenintheabsenceofbehavioraldata.

4.5 Explainability

Every recommendation includes a human-readable explanationgeneratedfromthedominantscoringsignal.For example: "Matches your recent interest in formal wear" (intent-based), "Popular with shoppers who have similar taste"(collaborative),or"Trendingrightnowinaccessories" (popularity-based).Thistransparencybuildsusertrustand enablesuserstorefinetheirpreferencesthroughfeedback.

5. VOICE CHATBOT SYSTEM

Theconversationalvoiceassistantprovidesanalternative shopping interface that enables hands-free product discoveryandtransactionmanagement.

5.1 Architecture

The chatbot system comprises five processing stages arrangedinapipeline:

1. Speech-to-Text:Voiceinputiscapturedthroughthe WebSpeechAPIonthefrontendandconvertedto text using either the browser's native speech recognitionortheOpenAIWhisperAPIforserversideprocessing.

2. Intent Detection and Entity Extraction:The text input is processed by GPT-4 with a structured promptthatclassifiestheuser'sintentintooneof 20definedcategories(productsearch,addtocart, orderstatus, recommendations, etc.)and extracts

relevant entities (product names, categories, quantities,priceranges).

3. Dialog State Management:A conversation state machine maintains context across multi-turn interactions,trackingthecurrenttopic,referenced products,andpendingactions.

4. ActionExecution:Basedonthedetectedintent,the chatbotinvokestheappropriatemicroserviceAPI productsearch,cartmanagement,orderqueries,or therecommendationengine.

5. Response Generation and Text-to-Speech:GPT-4 generates a natural language response incorporating the action results. The response is optionallyconvertedtoaudiousingtheElevenLabs text-to-speechAPIforvoiceoutput.

5.2 Supported Intents

Thechatbotsupportsthefollowingintentcategories:

Table 4: Chatbotintentcategorieswithexample utterances

Category Intents Example Utterance

Product Discovery Search,Details,Compare, Recommendations "Showmesummer dressesunderfifty dollars"

Shopping

Add to Cart, Remove, View Cart, Update Quantity "Add that blue dresstomycart"

Orders Checkout,Status,History, Cancel "Whatisthestatus ofmylastorder"

Navigation Browse Category, Filter Products

"Show me the accessories collection"

General Greeting, Help, FAQ, Goodbye "I need help findingagift"

5.3 Contexts-Aware Responses

Thechatbotintegrateswiththerecommendationengineto provide personalized product suggestions within the conversationalflow.Whenauserasksforrecommendations, thechatbotforwardstherequesttotheAIservicealongwith theconversationcontext,andtheresponseincludesproduct cardsrenderedinlinewithinthechatinterface.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 02 | Feb 2026 www.irjet.net p-ISSN: 2395-0072

6. RESULTS AND DISCUSSION

6.1

Performance Evaluation

The recommendation engine was evaluated on several performancemetricsundersimulatedloadconditions:

Table 5: Recommendationengineperformancemetrics

Metric

ResponseLatency <200ms ~45ms

MemoryUsage <500MB ~150MB

Throughput 100req/s 500+req/s

Cold-Start Handling Graceful degradation Verified

Thesub-50millisecondlatencyisachievedthroughtheperuser indexed query design, which avoids the overhead of full-dataset operations. The memory footprint remains under150MBbecausethesystemneverloadsthecomplete productcatalogoruser-itemmatrixintomemory.Instead, eachrecommendationrequesttriggersaseriesoftargeted databasequeriesthatretrieveonlythedatarelevanttothe specificuser.

6.2 Scalability Analysis

The scalability advantage of StyleWise over traditional recommendationapproachesissummarizedbelow:

Table 6: Comparisonoftraditionalandproposed approaches

Aspect Traditional Approach StyleWise Approach

MemoryModel Fulluser-itemmatrix inmemory Per-user indexed queries

Computational Complexity O(n × m) matrix operations O(k) per-user operations

ModelUpdates Full retraining required Real-time incremental updates

Cold-StartUsers Norecommendations Graceful fallback to trending

Aspect Traditional Approach StyleWise Approach generated products

Explainability Black-boxoutput Human-readable explanations

ResponseTime Seconds ~45milliseconds

6.3 System Integration

The Docker-containerized deployment enables seamless orchestrationofallplatformservices.TheDockerCompose configurationmanagesnineservices:MongoDB,Redis,API Gateway, Product Service, Order Service, AI RecommendationService,ChatbotService,AdminService, and the Next.js Frontend. Health checks ensure service dependenciesare met beforestartup,andvolumemounts providedatapersistenceacrosscontainerrestarts.

6.4 User Experience

Theonboardingflowsuccessfullyaddressesthecold-start problem by collecting four preference dimensions (style, budget, occasion, colors) during registration. Users who complete onboarding receive personalized recommendations from their first session, with recommendation relevance improving progressively as behavioraldataaccumulates.Thevoicechatbotprovidesan alternativeinteractionmodalitythatreducesthecognitive loadofproductdiscovery,particularlyeffectiveforhandsfreeandaccessibilityusecases.

7. CONCLUSIONS

This paper presented StyleWise, a comprehensive AIpowerede-commerceSaaSplatformthatintegratesahybrid recommendation engine with a conversational voice assistant. The two-stage recommendation architecture consisting of multi-source candidate generation and fivefactor hybrid ranking achieves a response latency of approximately45millisecondswhilemaintainingamemory footprint under 150 megabytes, demonstrating that intelligent personalization can be delivered without the computationaloverheadoftraditionalmatrixfactorization approaches.

The key contributions of this work are: (1) a scalable candidate generation strategy that combines category retrieval,contentsimilarity,localizedcollaborativefiltering, andtrendinganalysiswithoutrequiringfull-datasetloading; (2)ahybridscoringformulawithfiveweightedsignalsthat balancesshort-termintentwithlong-termpreferences;(3)a dynamic cold-start handling mechanism that shifts

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 02 | Feb 2026 www.irjet.net p-ISSN: 2395-0072

recommendation source weights based on the user's interaction maturity; (4) a GPT-4 powered voice chatbot with20definedintentsforconversationalproductdiscovery and shopping; and (5) a production-ready microservices architecture containerized with Docker for horizontal scalability.

Future work includes the integration of semantic embeddingsusingsentence-transformersforrichercontent similarity, implementation of A/B testing frameworks for continuous ranking weight optimization, deployment of neural collaborative filtering models for improved personalizationaccuracy,andintegrationofreal-timeevent streaming with Apache Kafka for higher-throughput behavioraldataprocessing.

ACKNOWLEDGEMENT

Theauthorsexpressgratitudeto[YourCollege/University Name] for providing the infrastructure and guidance necessary for the successful completion of this project. Special thanks to [Guide Name] for valuable mentorship throughouttheresearch.

REFERENCES

[1] M. J. Pazzani and D. Billsus, "Content-Based Recommendation Systems,"The Adaptive Web, Springer,2007,pp.325–341.

[2] P.Lops,M.deGemmis,andG.Semeraro,"ContentbasedRecommenderSystems:StateoftheArtand Trends,"Recommender Systems Handbook, Springer,2011,pp.73–105.

[3] J.S.Breese,D.Heckerman,andC.Kadie,"Empirical AnalysisofPredictiveAlgorithmsforCollaborative Filtering,"Proc.14thConf.UncertaintyinArtificial Intelligence,1998,pp.43–52.

[4] Y. Koren, R. Bell, and C. Volinsky, "Matrix Factorization Techniques for Recommender Systems,"IEEEComputer,vol.42,no.8,Aug.2009, pp.30–37,doi:10.1109/MC.2009.263.

[5] R. Burke, "Hybrid Recommender Systems: Survey andExperiments,"UserModelingandUser-Adapted Interaction,vol.12,no.4,2002,pp.331–370.

[6] R.Burke,"HybridWebRecommenderSystems,"The AdaptiveWeb,Springer,2007,pp.377–408.

[7] P.Covington,J.Adams,andE.Sargin,"DeepNeural Networks for YouTube Recommendations,"Proc. 10thACMConf.RecommenderSystems,2016,pp. 191–198,doi:10.1145/2959100.2959190.

[8] A. I. Schein, A. Popescul, L. H. Ungar, and D. M. Pennock, "Methods and Metrics for Cold-Start Recommendations,"Proc. 25th Annual Int. ACM SIGIRConf.,2002,pp.253–260.

[9] C. Conversica, "Conversational Commerce: The StateofAI-PoweredShopping,"JournalofRetailing andConsumerServices,vol.62,2021,pp.102–115.

[10] S. Reddy, "FastAPI: Modern Python Web Framework for Building APIs,"Python Software FoundationDocumentation,2023.

[11] T. Vercel, "Next.js: The React Framework for Production,"VercelDocumentation,2024.

[12] MongoDB Inc., "MongoDB Manual: DocumentOrientedDatabase,"MongoDBDocumentation,v7.0, 2024.