I. Introduction
E-commerce platforms let customers browse, compare and pay online, yet the customer still performs most of the work: translating a need into keywords, applying filters, comparing specifications and finding related items. Large language models (LLMs) can interpret free-form requests, which makes it possible to delegate part of this work to a software agent. Agentic AI, in which a system pursues a goal by making decisions and using tools rather than answering isolated commands, is an active research area1, and LLM-based agents combine planning, memory and tool interaction to complete multi-step tasks2. Approaches such as ReAct interleave reasoning with actions on external tools3.
Shopping is a demanding setting for such agents because the final step has financial consequences. An agent that misreads a budget, invents a product or charges a customer without consent is unacceptable, even if its recommendations are otherwise good. The central design question of this work is therefore how to let an LLM contribute flexible language understanding while keeping every money-related decision bounded, explainable and gated by the human.
This paper presents TechNest, an implemented and publicly deployed AI buyer and agentic commerce prototype. A customer types a request such as “laptop around 50000”; the system extracts intent, searches a merchant catalogue, ranks candidates, suggests a complementary product within the remaining budget, assembles a basket, and pauses for explicit approval. Only after approval is a Razorpay order created. The payment signature is verified on the server and the outcome is stored in an audit trail that drives a merchant dashboard. The contributions are deliberately scoped:
1) A division-of-labour design in which the LLM handles intent extraction, candidate selection and justification, while deterministic code enforces hard constraints, totals, verification and auditing (Section VI).
2) A stateful, interruptible workflow implemented in LangGraph, with a human approval gate and a payment-wait state that pause and resume a single session.
3) A server-side validation and verification path that rebuilds the approved basket from catalogue data and treats a payment as successful only after signature verification.
4) A transaction audit trail and merchant dashboard that record approvals, successes and failures.
5) An honest functional evaluation of the deployed system, with explicit identification of what has not yet been measured.
TechNest is a reference implementation on a fictional catalogue and in payment Test Mode; it does not claim to solve agentic commerce in general. The rest of the paper reviews related work, states the problem and gap, describes the architecture, methodology and implementation, reports the evaluation, and discusses limitations and future work.
II. Background and Related Work
A. AI in E-Commerce and Recommendation
Recommendation is the most established use of AI in online retail. Ko et al. survey recommendation models, covering content-based, collaborative and hybrid filtering and their application fields4. Conversational recommender systems extend this by letting users express and refine preferences over several dialogue turns; Gao et al. review their techniques and open challenges5. More recently, LLMs have been integrated into recommenders as a flexible natural-language interface, and Shehmir and Kashef survey this integration together with issues such as latency, bias and evaluation6. Explanations matter for trust: Said’s review of LLM-generated explanations found only six directly relevant studies among 232 search results and observes that LLMs mainly produce human-readable justifications that improve perceived transparency rather than exposing internal model mechanics7. These works concentrate on what to recommend; they do not address how a recommendation becomes a controlled purchase.
B. Agentic AI, LLM Agents and Tool Use
Acharya et al. characterise agentic AI by autonomy, goal orientation, adaptability, reasoning and interaction with the environment, contrast it with task-specific traditional AI, and discuss human-in-the-loop configurations in which operators approve certain actions8. Wang et al. propose a unified framework of four modules for LLM-based agents (profile, memory, planning and action), separating planning with and without feedback and distinguishing external tools from the model’s internal knowledge9. ReAct motivates environment feedback by prompting the model with thought–action–observation steps10. GPT4Tools shows that open-source LLMs can be taught to use tools through self-instruction, with negative samples so that the model learns when not to call a tool11. These works describe how agents plan and call tools; they leave open how a particular application should constrain the agent when the tool is a payment.
C. Grounded Shopping Agents and Their Evaluation
WebShop is a simulated shopping environment with 1.18 million real products and 12,087 crowd-sourced instructions in which an agent must search, inspect items, select options and buy a matching product12. Its best learned agent reached a success rate of about 29% compared with about 60% for expert humans, and choosing the correct product options was the largest source of failure13. The result suggests that leaving constraint handling entirely to a learned model is risky; TechNest consequently keeps category and budget in deterministic tools. WebShop also illustrates evaluation by task success and reward, a style that a future TechNest benchmark could follow.
Table I: Comparison of Related Approaches
|
Approach |
Main capability |
Limitation for purchasing |
|
Recommenders14 |
Suggest items from user/item data |
Weak on complex language goals; no purchase step |
|
Conversational RS15 |
Refine preferences by dialogue |
Focus on recommendation, not payment |
|
LLM-based RS16 |
Contextual understanding |
Latency, bias, evaluation |
|
LLM explanations17 |
Readable justifications |
Must stay grounded in product data |
|
LLM agents18 |
Planning and tool use |
Generic; no payment constraints |
|
WebShop agents19 |
Language-driven shop actions |
Low success on option selection |
|
TechNest |
Bounded, approved, verified, audited purchase |
Small catalogue; Test Mode |
D. Human Oversight, Transparency and Auditability
Human-in-the-loop machine learning incorporates human knowledge and feedback into automated processes20. For purchases, the human is best positioned as an approver of an irreversible action. Work on AI-driven real-time traffic management in communication networks likewise notes that transparent, explainable decisions are needed for regulatory compliance and operational trust21. That study addresses network traffic rather than commerce, so we use it only for this general requirement, which carries over to payment-related decisions and motivates TechNest’s approval gate and audit trail. Related work from the project supervisor’s group applies neural-network and federated reinforcement-learning methods to adaptive real-time traffic signal control22. These systems also lie outside commerce; we cite them only as context for placing learning-based components in real-time decision loops where reliability matters, the same concern that motivates bounding the LLM’s role in TechNest.
E. Positioning
Table I summarises the approaches above. Each addresses one part of the shopping experience. TechNest does not propose a new recommendation algorithm; its contribution lies in combining language-driven discovery, bounded basket construction, human approval, verified payment and auditing in one resumable workflow.
III. Research Problem and Objectives
A. Problem Statement
In conventional e-commerce, customers search, filter and compare items manually and find complementary products by exploring catalogue sections. Keyword search may miss the meaning of a request that combines category, features and budget. An AI agent can reduce this effort, but it must not select outside the customer’s constraints, invent products, or complete a purchase without consent. The problem addressed here is to design and implement an AI buyer that (i) understands natural-language requests, (ii) respects hard constraints, (iii) explains its choices, (iv) requires human approval before payment, and (v) leaves a verifiable record of every outcome.
B. Objectives
1) Interpret natural-language shopping requests into category, budget and requirements.
2) Search and rank catalogue products and suggest one complementary item within the remaining budget.
3) Build a basket and require explicit human approval before payment.
4) Integrate Razorpay (Test Mode) with server-side signature verification and failure handling.
5) Record transactions in an audit trail and expose business insights to the merchant.
6) Report what the deployed system demonstrably does and what remains unmeasured.
IV. Research Gap and Motivation
The reviewed literature offers strong components: recommendation models, LLM interfaces, agent frameworks and shopping benchmarks. What is less visible is an application-level treatment of the last mile: how an LLM-driven buyer hands over from reasoning to payment under explicit control. Table II lists the practical gaps TechNest targets. These gaps are drawn from the limitations of the surveyed approaches and the design of conventional storefronts, not from a measured comparison.
Table II: Identified Gaps and the TechNest Approach
|
Gap |
TechNest approach |
|
Manual search and filtering |
Natural-language AI Buyer |
|
Search separated from basket and payment |
One LangGraph workflow from request to audit |
|
Budget left to the user |
Budget-aware filtering and add-on selection in code |
|
Automated purchase without control |
Mandatory human approval before order creation |
|
Payment trusted from the browser |
Server-side basket rebuild and signature verification |
|
No record of AI-assisted purchases |
SQLite audit trail and merchant dashboard |
V. Proposed System Architecture
A. Overview
TechNest keeps a conventional storefront, so customers can still browse and use a regular cart, and adds an AI Buyer for natural-language shopping. Figure 1 shows the architecture. A React frontend communicates over HTTPS with a FastAPI backend. The backend exposes catalogue, AI Buyer, approval, payment and merchant endpoints; runs the LangGraph workflow; calls a Groq-hosted LLM for intent understanding, ranking and add-on selection; calls Razorpay for order creation and signature verification; and stores audit records in SQLite.
Figure 1: Overall system and deployment architecture of TechNest.
B. Components and Responsibilities
The system comprises seven modules: (1) the customer interface (store, cart, AI Buyer page, checkout); (2) a JSON product catalogue with category, price, tags and specifications; (3) the AI Buyer (intent extraction, search, ranking, upselling, basket); (4) the approval module (interrupt, validation, resume); (5) the payment module (order, checkout data, verification); (6) transaction auditing; and (7) the merchant dashboard, which reads audit records and computes summary insights. Frontend, API, workflow, tools and data layers are kept separate so each can be replaced independently.
C. Deployment
The frontend is hosted on Render. The backend runs on an AWS EC2 instance behind an Nginx reverse proxy with a Let’s Encrypt certificate and a DuckDNS domain; CORS is limited to the configured frontend origin. AI and payment features depend on the availability of the Groq and Razorpay services.
VI. Methodology and Agentic Workflow
A. Workflow as a Stateful Graph
The AI Buyer is a LangGraph state graph whose nodes read and update one shared typed state (user input, extracted intent, candidates, selected and recommended products, basket, total, approval, Razorpay identifiers and payment status). Figure 2 shows the twelve nodes: understand_intent, search_catalog, Rank_products, upsell_crosssell_products, Build_basket, hitlnode, create_razorpay_order, prepare_checkout, payment_wait_node, verify_payment, handle_failure and audit_transaction. Conditional edges route the run: the graph ends if no product satisfies the constraints or if the customer rejects the basket, and after verification a success or failure route leads to the audit node.
Figure 2: Agentic workflow of the TechNest AI Buyer with conditional routes and two interrupts. Blue: LLM-assisted nodes; orange: human-gated or externally resumed pauses; grey: deterministic code.
B. Division of Work Between the LLM and Code
The key design decision is to let the LLM handle language and keep decisions with financial consequences in deterministic code (Table III). This is motivated by evidence that constraint and option selection is hard for learned shopping agents23. Each LLM call uses structured output with a fixed schema (intent, ranking and upsell schemas). The reply is validated and mapped back to catalogue records by product identifier, so the model cannot introduce a product that is not in the catalogue. In the vocabulary of24, planning is a fixed graph with conditional edges rather than free-form LLM planning; feedback comes from tool results and from the human decision; memory is the shared workflow state of the current run, and no long-term user memory is kept.
Table III: Responsibility Split Between the LLM and Deterministic Code
|
Task |
By |
Note |
|
Extract category, budget, requirements |
LLM |
Structured intent schema |
|
Normalise category wording |
Code |
Alias table to catalogue categories |
|
Category and budget filtering |
Code |
Hard constraints in search tool |
|
Choose best candidate |
LLM |
Only from filtered set, by product ID |
|
Choose add-on |
LLM |
Only within remaining budget; may return none |
|
Basket total |
Code |
Sum of server-side prices |
|
Approval validation |
Code |
Session, products, main item, total |
|
Payment verification |
Razorpay utility |
HMAC-based signature check |
|
Audit record |
Code |
SQLite insert |
C. Constraint Formulation
Let P be the catalogue, c the normalised category and B the extracted maximum budget (when the request states one). The search tool returns the candidate set
C = { p ∈ P : cat(p) = c ∧ price(p) ≤ B }. (1)
Preferences such as RAM, battery or camera are not hard filters; they are left to the ranking step, which selects one ps ∈ C. If C = ∅ the run ends rather than recommending an unrelated item. With the remaining budget R = B − price(ps), the add-on candidates are
U = { p ∈ P : cat(p) ≠ c ∧ price(p) ≤ R }, (2)
from which the LLM may select one item pu or none. The basket total is computed in code as
T = price(ps) + price(pu), (3)
with the second term omitted if no add-on is chosen. The LLM never computes T.
D. Human Approval and Payment Execution
The workflow is not a single request–response call. The first call to /api/buy (or its streaming variant) runs the graph until the approval interrupt and returns the proposed basket with a session identifier. The decision is sent to /api/approval, which validates it and resumes the same session; the graph then creates the order and pauses again in the payment-wait node. After checkout, the result is sent to /api/payment/verify or /api/payment/failure and the graph resumes, verifies the signature and writes the audit record (Fig. 3). State is stored by a checkpointer under the session identifier.
Figure 3: Human approval and payment execution workflow with the two interrupt points (sequence view).
E. Approval Validation
The approval endpoint does not trust the browser. It checks that the session exists and the decision is valid, that submitted products belong to the catalogue, and that the AI-selected main product is still in the basket. It then rebuilds the basket from server-side product data and recalculates the total before resuming the graph. The customer may remove the optional add-on, but a manipulated browser request cannot change the price.
F. Payment Verification, Failure Handling and Auditing
On approval the backend creates a Razorpay order with the amount in paise, and the frontend opens the hosted checkout. The browser then returns the payment, order and signature identifiers, and the verification node calls Razorpay’s signature utility. Success is recorded only if the check passes; an invalid signature, an exception or a failure event from the checkout sets the status to failed. In both cases the audit node inserts a record (request, selected and recommended products, total, approval decision, order identifier and payment status), so failed attempts are visible to the merchant and are not counted as revenue.
VII. System Implementation and Technologies
A. Technology Stack
Table IV summarises the stack. The catalogue search is a LangChain tool; LangGraph provides conditional routing, interrupts and an in-memory checkpointer; the LLM is accessed through the Groq API.
Table IV: Technology Stack of TechNest
|
Layer |
Technology |
Role |
|
Frontend |
React, JSX, Vite |
Views, cart state, API calls |
|
Backend |
Python, FastAPI, Uvicorn |
REST endpoints, ASGI server |
|
LLM |
Groq (openai/gpt-oss-20b) |
Intent, ranking, add-on |
|
Orchestration |
LangChain, LangGraph |
Prompts, structured output, stateful graph |
|
Payment |
Razorpay Test Mode |
Orders, checkout, verification |
|
Storage |
JSON, SQLite |
Catalogue, audit records |
|
Hosting |
Render, AWS EC2, Nginx, DuckDNS |
Frontend, backend, HTTPS, domain |
B. Catalogue and Intent Handling
The catalogue is a fixed JSON file of 14 fictional products in six categories (laptops, headphones, keyboards, mice, monitors and smartphones), each with identifier, name, brand, category, price, description, tags and a specification object. Customers use different words for the same category, so the extracted category is normalised through an alias table (for example, “phone” or “mobile” to “smartphone”) to one of the exact catalogue categories, which makes the subsequent search a deterministic comparison.
C. Ranking and Upselling Prompts
The ranking prompt tells the model to consider specifications such as RAM, processor, battery, display, storage and price and to choose only from the supplied candidates. The upsell prompt tells it to recommend a product only if it is genuinely useful with the selected product, never to recommend the selected item or invent products, and to return null otherwise. Structured output returns an identifier and a short explanation, and the node looks up the full record by identifier.
D. Interface and Dashboard
The frontend is a single-page application with views for the store, the AI Buyer, the cart and the merchant dashboard. The AI Buyer page streams workflow progress so the user sees the current stage while the LLM and tools run. The dashboard endpoint reads the SQLite audit_logs table and reports total transactions, successful and failed payments, revenue, average order value, top product, top category and the number of transactions with an AI upsell, together with a table of recent AI Buyer transactions. SQLite is used for auditing only and is not an inventory or order-management database.
E. Security Measures
API keys for Groq and Razorpay are read from server environment variables, and only the public Razorpay key identifier reaches the browser. Baskets and totals are recomputed on the server; payments are accepted only after signature verification; HTTPS is terminated at Nginx; and LLM output is constrained by schemas and mapped to catalogue identifiers, which limits the effect of incorrect or injected text in a request. Machine-learning-based malware detection25 is a complementary system-level defence and is not implemented in TechNest.
VIII. Evaluation, Results and Discussion
A. Evaluation Approach
Evaluation was carried out on the deployed application with Razorpay in Test Mode. It was functional: unit-level checks (catalogue loading, filters, signature verification, database initialisation), integration checks (workflow with catalogue and payment) and system-level checks (a complete purchase through the browser). It was not a quantitative performance study. We did not record latency, recommendation accuracy, conversion or user-study data, and none are claimed. Inspired by task-success evaluation in WebShop26, the observable outcomes are whether category and budget were applied, whether the add-on fits the remaining budget, whether approval preceded payment, and whether payment and audit status are correct.
Table V: Summary of Functional Test Cases
|
Group |
Focus |
Passed |
Pending |
|
TC (9) |
Storefront, cart, AI request, recommendation, add-on, basket total; unknown category, streaming |
7 |
2 |
|
HA (3) |
Approve, reject, tampered basket |
1 |
2 |
|
PY (5) |
Order, checkout, valid and failed payment, invalid signature |
4 |
1 |
|
DB (2) |
Audit record, dashboard |
2 |
0 |
|
BG (2) |
Budget filter, no affordable add-on |
0 |
2 |
|
Total (21) |
|
14 |
7 |
B. Functional Test Results
Twenty-one test cases were defined, grouped in Table V. Fourteen were exercised and are supported by screenshots in the project report; seven are defined but not yet recorded and are therefore reported as pending rather than passed.
The pending cases include several of the safety properties emphasised in the design (rejection, tampered basket, invalid signature, budget exclusion). Their implementation exists, but this paper does not claim empirical confirmation until they are recorded.
C. Sample Run
For the request “laptop around 50000” (Fig. 4), the intent was normalised to the laptop category with a budget of about INR 50,000. Of the three laptops in the catalogue, the TechNest LiteBook 15 (INR 44,999) was selected with a model-generated justification noting a balanced combination of performance, storage, display and battery at an affordable price. A wireless keyboard, the TechNest Compact K2 (INR 1,799), was suggested as an optional add-on, and the basket total of INR 46,798 equals the sum of the two prices. The interface states that nothing will be charged until the customer approves, and offers Reject and Approve actions.
Figure 4: AI Buyer justification, optional add-on and basket awaiting human approval (cropped screenshot).
D. Payment and Audit Observations
The deployed system completed the full flow from request to verified payment in Razorpay Test Mode (Fig. 5). Table VI shows the dashboard summary recorded during testing: eight transactions, six successful and two failed. The two failed payments (a headphones and a smartphone request) were stored and displayed rather than counted as revenue, which demonstrates the failure path.
Figure 5: Razorpay checkout in Test Mode: (a) payment options for a basket of INR 54,999; (b) successful test payment confirmation.
|
(a) |
(b) |
Table VI: Dashboard Summary Recorded During Testing
|
Metric |
Value |
|
Total transactions |
8 |
|
Successful / failed payments |
6 / 2 |
|
Revenue (successful payments) |
INR 345,992 |
|
Average order value |
INR 57,665.33 |
|
Top product / top category |
TechNest UltraBook X / Laptop |
|
Transactions with AI upsell |
4 |
E. Discussion
Four observations follow from the recorded activity (Fig. 6). First, requests phrased as “under 3 k”, “under 40k” and “2 lakh rupees” were converted into budget-filtered purchases, which suggests an advantage of LLM-based intent extraction over fixed keyword rules; this is an observation from a handful of requests, not a measured accuracy. Second, the justification and add-on were shown before any payment. Third, failed payments were recorded and not treated as sales. Fourth, the approval step with server-side recalculation keeps the final amount under server control.
Figure 6: Merchant dashboard with business insights and audit trail.
The evidence supports the claim that the workflow behaves as designed on the exercised paths. It does not support claims about recommendation quality, latency, user preference or business impact. The dashboard figures come from a small number of test transactions and should not be read as performance results. The LLM-driven ranking is a reasoning step over a 14-product catalogue; the system includes no separately trained predictive model, such as a purchase-likelihood classifier.
IX. Limitations and Future Scope
A. Limitations
The catalogue is a fictional 14-product JSON file, so recommendation quality cannot be generalised to real marketplaces, and no live inventory or marketplace integration exists. Razorpay Test Mode does not prove real settlement. The LLM is an external API, so latency and availability vary, and its ranking may not be optimal for every customer. Workflow state is held in memory, so a paused run is lost if the server restarts. There are no user accounts, and the dashboard shows all recorded transactions. SQLite auditing is suitable for demonstration rather than production scale. Finally, seven test cases remain unrecorded and no quantitative metrics were collected.
B. Future Scope
The most valuable next step is a benchmark of shopping requests with expected categories, budgets and acceptable products, scored with task-success measures like those in WebShop27, together with latency and cost measurement and completion of the pending tests. Further directions include: splitting the single graph into cooperating specialised agents, an architecture discussed for agentic AI28; consented personalisation and multi-turn clarification29; privacy-preserving personalisation, where protocols for retrieving private data from vertically partitioned data30 and federated learning31 may inform approaches that avoid centralising customer data; feature-level evidence alongside LLM justifications and user studies of their clarity32; live catalogue and inventory integration; persistent checkpoints, authentication and role-based merchant access; and, as a separate component, a trained purchase-likelihood model, which TechNest does not currently implement.
X. Conclusion
TechNest shows how an LLM-driven buyer can be embedded in a commerce workflow without surrendering control of the money action. By confining the LLM to language understanding, candidate selection and justification, enforcing constraints and totals in code, pausing for explicit human approval, verifying payment signatures on the server and auditing every outcome, the system turns a natural-language request into a reviewable, verifiable purchase. The deployed prototype has clear limits: a small fictional catalogue, Test Mode payments, in-memory state and a functional rather than quantitative evaluation. Within that scope, it provides a concrete and inspectable pattern for approval-gated agentic commerce and a foundation for the benchmarked, production-oriented work outlined above.
Endnotes
1. D. B. Acharya, K. Kuppan, and B. Divya, “Agentic AI: Autonomous intelligence for complex goals—A comprehensive survey,” IEEE Access, vol. 13, pp. 18912–18936, 2025, doi: 10.1109/ACCESS.2025.3532853.
2. L. Wang et al., “A survey on large language model based autonomous agents,” Frontiers of Computer Science, vol. 18, no. 6, Art. no. 186345, 2024, doi: 10.1007/s11704-024-40231-1.
3. S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, “ReAct: Synergizing reasoning and acting in language models,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2023. [Online]. Available: https://arxiv.org/abs/2210.03629
4. H. Ko, S. Lee, Y. Park, and A. Choi, “A survey of recommendation systems: Recommendation models, techniques, and application fields,” Electronics, vol. 11, no. 1, Art. no. 141, 2022, doi: 10.3390/electronics11010141.
5. C. Gao, W. Lei, X. He, M. de Rijke, and T.-S. Chua, “Advances and challenges in conversational recommender systems: A survey,” AI Open, vol. 2, pp. 100–126, 2021, doi: 10.1016/j.aiopen.2021.06.002.
6. S. Shehmir and R. Kashef, “LLM4Rec: A comprehensive survey on the integration of large language models in recommender systems—Approaches, applications and challenges,” Future Internet, vol. 17, no. 6, Art. no. 252, 2025, doi: 10.3390/fi17060252.
7. A. Said, “On explaining recommendations with large language models: A review,” Frontiers in Big Data, vol. 7, Art. no. 1505284, 2025, doi: 10.3389/fdata.2024.1505284.
8. D. B. Acharya, K. Kuppan, and B. Divya, “Agentic AI: Autonomous intelligence for complex goals—A comprehensive survey,” IEEE Access, vol. 13, pp. 18912–18936, 2025, doi: 10.1109/ACCESS.2025.3532853.
9. L. Wang et al., “A survey on large language model based autonomous agents,” Frontiers of Computer Science, vol. 18, no. 6, Art. no. 186345, 2024, doi: 10.1007/s11704-024-40231-1.
10. S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, “ReAct: Synergizing reasoning and acting in language models,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2023. [Online]. Available: https://arxiv.org/abs/2210.03629
11. R. Yang, L. Song, Y. Li, S. Zhao, Y. Ge, X. Li, and Y. Shan, “GPT4Tools: Teaching large language model to use tools via self-instruction,” in Advances in Neural Information Processing Systems (NeurIPS), 2023.
12. S. Yao, H. Chen, J. Yang, and K. Narasimhan, “WebShop: Towards scalable real-world web interaction with grounded language agents,” in Advances in Neural Information Processing Systems (NeurIPS), 2022.
13. S. Yao, H. Chen, J. Yang, and K. Narasimhan, “WebShop: Towards scalable real-world web interaction with grounded language agents,” in Advances in Neural Information Processing Systems (NeurIPS), 2022.
14. H. Ko, S. Lee, Y. Park, and A. Choi, “A survey of recommendation systems: Recommendation models, techniques, and application fields,” Electronics, vol. 11, no. 1, Art. no. 141, 2022, doi: 10.3390/electronics11010141.
15. C. Gao, W. Lei, X. He, M. de Rijke, and T.-S. Chua, “Advances and challenges in conversational recommender systems: A survey,” AI Open, vol. 2, pp. 100–126, 2021, doi: 10.1016/j.aiopen.2021.06.002.
16. S. Shehmir and R. Kashef, “LLM4Rec: A comprehensive survey on the integration of large language models in recommender systems—Approaches, applications and challenges,” Future Internet, vol. 17, no. 6, Art. no. 252, 2025, doi: 10.3390/fi17060252.
17. A. Said, “On explaining recommendations with large language models: A review,” Frontiers in Big Data, vol. 7, Art. no. 1505284, 2025, doi: 10.3389/fdata.2024.1505284.
18. L. Wang et al., “A survey on large language model based autonomous agents,” Frontiers of Computer Science, vol. 18, no. 6, Art. no. 186345, 2024, doi: 10.1007/s11704-024-40231-1; S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, “ReAct: Synergizing reasoning and acting in language models,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2023. [Online]. Available: https://arxiv.org/abs/2210.03629; R. Yang, L. Song, Y. Li, S. Zhao, Y. Ge, X. Li, and Y. Shan, “GPT4Tools: Teaching large language model to use tools via self-instruction,” in Advances in Neural Information Processing Systems (NeurIPS), 2023.
19. S. Yao, H. Chen, J. Yang, and K. Narasimhan, “WebShop: Towards scalable real-world web interaction with grounded language agents,” in Advances in Neural Information Processing Systems (NeurIPS), 2022.
20. E. Mosqueira-Rey, E. Hernández-Pereira, D. Alonso-Ríos, J. Bobes-Bascarán, and Á. Fernández-Leal, “Human-in-the-loop machine learning: A state of the art,” Artificial Intelligence Review, vol. 56, pp. 3005–3054, 2023, doi: 10.1007/s10462-022-10246-w.
21. G. V. Sonawane, P. R. Patil, and P. R. Bhaladhare, “AI for real-time traffic management in communication networks,” Metallurgical and Materials Engineering, 2025.
22. G. V. Sonawane, A. Y. Panpatil, R. R. Raskar, M. B. Sonje, E. Muniyandy, et al., “Topology-aware federated graph reinforcement learning for real-time edge-based adaptive traffic signal control in smart cities,” in Proc. 2026 Int. Conf. [VENUE TITLE — complete before submission], 2026; G. V. Sonawane, P. R. Patil, and P. R. Bhaladhare, “Neural network-based traffic control system,” Journal of Electrical Systems, vol. 20, no. 3, 2024, doi: 10.52783/jes.7946.
23. S. Yao, H. Chen, J. Yang, and K. Narasimhan, “WebShop: Towards scalable real-world web interaction with grounded language agents,” in Advances in Neural Information Processing Systems (NeurIPS), 2022.
24. L. Wang et al., “A survey on large language model based autonomous agents,” Frontiers of Computer Science, vol. 18, no. 6, Art. no. 186345, 2024, doi: 10.1007/s11704-024-40231-1.
25. G. V. Sonawane, “Effective technique used for malware detection using machine learning,” International Research Journal of Engineering and Technology, vol. 6, no. 12, pp. 1102–1103, 2019.
26. S. Yao, H. Chen, J. Yang, and K. Narasimhan, “WebShop: Towards scalable real-world web interaction with grounded language agents,” in Advances in Neural Information Processing Systems (NeurIPS), 2022.
27. S. Yao, H. Chen, J. Yang, and K. Narasimhan, “WebShop: Towards scalable real-world web interaction with grounded language agents,” in Advances in Neural Information Processing Systems (NeurIPS), 2022.
28. D. B. Acharya, K. Kuppan, and B. Divya, “Agentic AI: Autonomous intelligence for complex goals—A comprehensive survey,” IEEE Access, vol. 13, pp. 18912–18936, 2025, doi: 10.1109/ACCESS.2025.3532853; L. Wang et al., “A survey on large language model based autonomous agents,” Frontiers of Computer Science, vol. 18, no. 6, Art. no. 186345, 2024, doi: 10.1007/s11704-024-40231-1.
29. H. Ko, S. Lee, Y. Park, and A. Choi, “A survey of recommendation systems: Recommendation models, techniques, and application fields,” Electronics, vol. 11, no. 1, Art. no. 141, 2022, doi: 10.3390/electronics11010141; C. Gao, W. Lei, X. He, M. de Rijke, and T.-S. Chua, “Advances and challenges in conversational recommender systems: A survey,” AI Open, vol. 2, pp. 100–126, 2021, doi: 10.1016/j.aiopen.2021.06.002.
30. G. V. Sonawane and N. Thoutam, “Design and analysis of three party protocol to retrieve private data by using vertical partitioned data,” International Journal of Electronics, Communication and Soft Computing, 2015.
31. G. V. Sonawane, A. Y. Panpatil, R. R. Raskar, M. B. Sonje, E. Muniyandy, et al., “Topology-aware federated graph reinforcement learning for real-time edge-based adaptive traffic signal control in smart cities,” in Proc. 2026 Int. Conf. [VENUE TITLE — complete before submission], 2026.
32. A. Said, “On explaining recommendations with large language models: A review,” Frontiers in Big Data, vol. 7, Art. no. 1505284, 2025, doi: 10.3389/fdata.2024.1505284.