Silver Spectrum TechFest 2025

A National Level 48-Hour Hackathon - 28, 29, and 30 October 2025

Problem Statements to be given to registered participants

Domain PS Code Problem Statement

NH-SJC-PS-001 Autonomous supply‑chain optimiser (agentic RAG)

Gen AI for Business

NH-SJC-PS-002 Agentic compliance assistant for finance
Applications

NH-SJC-PS-003 Cross‑functional marketing & finance campaign planner

Phishing email detector, Threat intelligence summarizer, Log

NH-SJC-PS-004 analysis + incident triage, Security policy assistant

Cyber Security
NH-SJC-PS-005 Spot the Insider Threat

NH-SJC-PS-006 AI-Powered Cyber Threat Detection Dashboard

Clinical triage assistant, Medical literature summarizer,

NH-SJC-PS-007 Medication adherence coach, Hospital resource scheduler,
Radiology report helper
Artificial Intelligence /
Machine Learning / AI teaching assistant that uses the Socratic method to teach
Deep Learning NH-SJC-PS-008 sorting
algorithms

NH-SJC-PS-009 Banking Fraud Detection using Multimodal Transactions

NH-SJC-PS-010 AI-Powered IoT Coaching System for Sports Training

Internet of Things
NH-SJC-PS-011 Smart Home Intrusion Detection using Multimodal Sensors
(IoT)

NH-SJC-PS-012 Defect Detection in Manufacturing using Multimodal Sensors

Innovation using Web

NH-SJC-PS-013 Front-Running Attack Detection & Mitigation System
3.0

Page 1 of 13
Domain-1
Gen AI for Business Applications
NH-SJC-PS-001. Autonomous supply‑chain optimiser (agentic RAG)
Retailers manage huge volumes of purchasing, inventory, logistics and demand‑forecasting data.
Inspired by AIMultiple’s description of agentic RAG, where an AI agent autonomously routes queries
to appropriate databases and self‑corrects, teams should build a multi‑agent system that accepts
natural‑language questions from planners (e.g., “Which warehouse should ship March orders to our
eastern region to minimise cost and meet service‑level agreements?”). The system must:
 Use a routing agent to decide which data sources to query (inventory DB, ERP orders table,
shipping‑cost API, etc.), generate appropriate SQL/API calls and retrieve the data. The agent
should perform multiple function calls and refine its approach if initial results are insufficient.
 Coordinate with a reasoning agent to compute optimal decisions—such as order–warehouse
assignments or route schedules—and provide interpretable explanations.
 Present a concise report showing the data sources used, the reasoning steps and the final
recommendation.
This task demonstrates how agentic RAG can enable autonomous, multi‑turn information retrieval and
decision‑making in supply‑chain management.
NH-SJC-PS-002. Agentic compliance assistant for finance
Financial institutions face complex regulations. A standard chatbot may not access internal policies,
whereas RAG allows a model to pull in enterprise‑specific data and produce accurate outputs.
Participants will build an agentic assistant that:
 Ingests regulatory documents, internal risk policies and recent compliance rulings. When a user
asks a question—such as “How do the latest Reserve Bank of India guidelines affect our
mortgage‑underwriting criteria?”—a reasoning agent autonomously retrieves and synthesises
relevant information from multiple sources and delivers a response with citations.
 Highlights ambiguities or conflicting regulations, flags potential risks and suggests follow‑up
questions.
 Demonstrates agentic features—autonomous decision‑making and real‑time adaptation—by
refining its responses based on user feedback and evolving regulations.
This challenge underscores how agentic AI can augment compliance teams by offering context‑aware
answers and risk insights.
NH-SJC-PS-003. Cross‑functional marketing & finance campaign planner
The CIO report notes that AI agents automate repetitive tasks and adapt to changing conditions across
domains like customer service, supply chain and IT operations. Planning a marketing campaign often
requires coordination between marketing, finance and inventory teams. Participants will create a
multi‑agent platform that:
 Uses RAG to access CRM data (customer profiles), finance budgets and real‑time inventory
levels.
 Plans a product promotion by generating targeted customer segments, projecting revenue,
checking budget constraints and suggesting discount levels.
 Deploys multiple agents: one agent proposes creative content, another agent validates financial
feasibility, and a third checks inventory availability. Agents collaborate and negotiate, adjusting
the plan autonomously.
 Presents the final campaign plan along with supporting data and explanations, adapting when
underlying data changes.

Page 2 of 13
Domain-2
Cyber Security
NH-SJC-PS-004. Phishing email detector, Threat intelligence summarizer, Log
analysis + incident triage, Security policy assistant
Pick a cybersecurity use case
 Phishing email detector + auto-responder: a GenAI agent analyzes incoming emails, flags phishing
attempts, drafts safe responses.
 Threat intelligence summarizer: an agent retrieves the latest CVEs/threat feeds and summarizes
risks for security teams.
 Log analysis + incident triage: ingest system logs, use RAG to identify anomalies, and recommend
mitigation steps.
 Security policy assistant: generate or validate compliance policies automatically.
 User-friendly security coach: explain security alerts or incidents to non-technical users in plain
language.
Build the GenAI workflow
o Use a GenAI framework such as LangChain, LangGraph, LlamaIndex, AutoGen, etc.
o Show how your system retrieves, analyzes, and generates insights or actions.
o Integrate at least one external data source or tool (threat feeds, logs, APIs, databases).
o Optionally, use RAG (Retrieval-Augmented Generation) or Agentic workflows (multiple
cooperating agents for retrieval, reasoning, and action).
Host/Run it
o Deploy on AWS or Azure (Bedrock/OpenAI on Azure, Lambda/Functions,
DynamoDB/CosmosDB, etc.) or run locally / cloud-agnostic.
Prepare a short demo
o 5–7 minute demo showing how the system works
o 1–2 page architecture diagram/slide deck explaining agents, tools, and state flow
Technologies & Resources
o GenAI frameworks: LangChain, LangGraph, LlamaIndex, Microsoft AutoGen, Hugging Face
Transformers
o Data sources: CVE/NVD feeds, VirusTotal, Shodan, Syslog/Kibana, AWS GuardDuty, Azure
Sentinel
o Cloud (optional): AWS (Bedrock, Lambda, DynamoDB, S3), Azure (OpenAI, Functions,
CosmosDB, Sentinel), or on-prem [Link] open APIs or security datasets
Tips
o Use RAG to enrich LLM responses with up-to-date threat intelligence.
o Design agentic workflows where one agent retrieves threat data, another analyzes, another drafts
actions/reports.
o Focus on privacy & security in your own implementation (no leaking of credentials).
o Show how your GenAI solution would fit into a real security operations workflow.

Page 3 of 13
NH-SJC-PS-005. “Spot the Insider Threat”

Background:
In IT companies, thousands of employees access emails, cloud apps, code repositories, and sensitive
client data daily. While firewalls and antivirus protect against outsiders, some of the biggest risks come
from inside — like an employee accidentally uploading data to public drives, sharing passwords, or
using personal USB devices.

The Challenge:
Design a simple solution to detect and alert when there is a potential insider threat.

Your solution should:

 Monitor basic activities such as unusual file uploads, bulk downloads, or logins from unexpected
locations.
 Show real-time alerts like “Large data download detected from HR folder at 2 AM.”
 Keep employee privacy protected (avoid storing personal details unnecessarily).
 Provide a clean dashboard with a simple risk score: Low / Medium / High. [a simple
dashboard/web page]
 Simulate 3–4 basic anomalies: like login at odd hours, large file download, or multiple failed
login attempts.
 Build a basic rule-based detector (no need for advanced ML).

Bonus:
 Add a feature to suggest quick actions: “Suspend suspicious session” or “Notify IT Security.”
 Show how the tool scales for a company with 1,000+ employees.

Goal:
Help IT companies catch risky activities early and prevent data leaks — keeping both client data and
company reputation safe.

Page 4 of 13
NH-SJC-PS-006. AI-Powered Cyber Threat Detection Dashboard
Cybersecurity – Focused on proactive threat detection, visualization, and simulation using AI.
Modalities:
 Text: Threat logs, system alerts, and incident reports.
 Image: Screenshots of attack patterns, phishing attempts, malware signatures.
 Video: Security camera feeds or recorded attack simulations for anomaly detection.

Detailed Steps to Implement:

1. Data Collection &Preprocessing:

 Aggregate threat logs from firewalls, IDS/IPS systems, and SIEM tools.
 Normalize and clean data for consistency across sources.
2. Threat Pattern Analysis:

 Use NLP models to extract insights from textual logs (e.g., suspicious IPs, attack vectors).
 Apply computer vision techniques to detect anomalies in images or video feeds.
3. Generative Modeling for Synthetic Scenarios:

 Implement GANs or diffusion models to create synthetic attack scenarios for training and stress-
testing.
 Simulate phishing campaigns, DDoS attacks, and ransomware behaviors in a controlled
environment.
4. Dashboard Development:

 Build an interactive dashboard using frameworks like Streamlit, Dash, or React.

 Include real-time threat visualization (heatmaps, timelines, attack graphs).
 Provide drill-down capabilities for detailed analysis of each threat.
5. Alerting & Recommendations:
 Integrate AI-driven alerts for high-risk patterns.
 Suggest mitigation strategies based on historical data and predictive analytics.
Expected Output:
A fully interactive dashboard that:

 Displays real-time threat visualizations (maps, graphs, timelines).

 Offers synthetic attack simulations for training and preparedness.
 Provides actionable insights and recommendations for security teams.

Page 5 of 13
Domain-3
Artificial Intelligence / Machine Learning / Deep Learning

NH-SJC-PS-007. Medical Assistant

Pick a medical use case
 Examples:
o Clinical triage assistant: intake patient symptoms, retrieve guidelines, generate preliminary triage
notes.
o Medical literature summarizer: fetch and summarize latest research for doctors.
o Medication adherence coach: personalized reminders + explanations for patients.
o Hospital resource scheduler: agentic system that retrieves resource data, forecasts demand, and
schedules staff.
o Radiology report helper: retrieve historical cases, help draft preliminary reports.
Build the GenAI workflow
 Use LangChain or LangGraph to define agents, tools, memory/state, and workflows.
 Integrate RAG or Agentic RAG to pull from medical guidelines, research papers, or hospital data.
 Show state propagation between nodes/agents.
 Include at least one external data source or service (e.g., PubMed API, FHIR server, hospital
database).
Host/Run it
 Deploy on AWS or Azure (Bedrock/OpenAI on Azure, Lambda/Functions,
DynamoDB/CosmosDB, HealthLake, Azure Health Data Services) or run locally/cloud-agnostic.
Prepare a short demo
 5–7 minute demo showing how the system works
 1–2 page architecture diagram/slide deck explaining agents, tools, RAG components, and state flow
Technologies & Resources
• GenAI frameworks: LangChain, LangGraph, LlamaIndex, Microsoft AutoGen, Hugging Face
Transformers
• Medical data sources: PubMed, WHO guidelines, FHIR servers, open-source EHR data sets (e.g.
MIMIC-III), hospital APIs (synthetic/test only)
• Cloud (optional): AWS (Bedrock, HealthLake, Lambda, DynamoDB, S3), Azure (OpenAI, Health
Data Services, Functions, CosmosDB)
Tips
• Use RAG to enrich LLM responses with up-to-date medical knowledge.
• Agentic workflows: design one agent for retrieval (guidelines, literature), another for reasoning/answer
generation, and others for acting or orchestrating tasks.
• Focus on privacy & security of patient data — only use de-identified or synthetic data for the
hackathon.
• Show how your workflow could realistically fit into a clinical or hospital workflow.

Page 6 of 13
[Link] an AI teaching assistant that uses the Socratic Method to
teach sorting algorithms.
The assistant should:
1. Ask probing questions instead of revealing answers directly
2. Adapt follow-up questions based on student responses
3. Guide students to discover answers through their own reasoning
4. Maintain focus on sorting algorithms
5. Handle the computational complexity efficiently
Target Topic: Data Structures and Algorithms - Sorting Algorithms
This topic is chosen because it is fundamental to computer science education, familiar to most software
engineering students, and well-defined in scope, making it ideal for a focused prototype.
Example 1: Time Complexity Query
Student Question: "What is the time complexity of bubble sort?"
Traditional Approach (NOT what we want): "The time complexity of bubble sort is O(n²)."
Socratic Approach (What your solution should do):
Step 1: "Let's think about how each algorithm works. Can you describe how merge sort divides the
problem?"
Step 2: "Good! Now, how does that division strategy affect the number of operations needed?"
Step 3: "Now compare that with bubble sort - how many comparisons does it make for an array of size
n?"
Step 4: Guide the student to compare O(n log n) vs O(n²) on their own
Example 2: When/Why to Use an Algorithm
Student Question: "When should I use quick sort?"
Traditional Approach (NOT what we want): "Quick sort is good for large datasets and has an average
time complexity of O(n log n), but worst case is O(n²) when the pivot is poorly chosen."
Socratic Approach (What your solution should do):
Step 1: "What factors do you think are important when choosing a sorting algorithm for a project?"
Step 2: "Interesting! How do you think quick sort performs on average compared to its worst case?"
Step 3: "What might cause quick sort to hit its worst-case scenario?"
Step 4: "Given what you've discovered, in what situations would quick sort be a good choice?"
Example 3: Implementation Details
Student Question: "How does merge sort work?"
Traditional Approach (NOT what we want): "Merge sort divides the array into halves recursively until
single elements remain, then merges them back in sorted order."
Socratic Approach (What your solution should do):
Page 7 of 13
Step 1: "Have you heard of the divide-and-conquer strategy in algorithms?"
Step 2: "How do you think we could apply that strategy to sorting?"
Step 3: "If we keep dividing an array in half, what's the smallest piece we can get?"
Step 4: "Great! Now, if you have two sorted arrays, how would you combine them into one sorted
array?"
Step 5: Guide the student to piece together the full merge sort process
Example 4: Handling Incorrect Answers
Student Question: "Is bubble sort O(n) time complexity?"
Traditional Approach (NOT what we want): "No, that's incorrect. Bubble sort is O(n²)."
Socratic Approach (What your solution should do):
Step 1: "Let's think about this together. In bubble sort, how many times do we need to pass through the
array?"
Step 2: "And in each pass, how many comparisons might we make?"
Step 3: "So if we have n elements, and we make up to n passes, with up to n comparisons each pass,
what does that tell us?"
Step 4: "Right! So would you like to reconsider your answer about the time complexity?"
FINAL NOTE
Remember: The goal is not just to build an AI that knows the answers, but one that teaches students how
to find them. Your assistant should make students think, not just provide information.

NH-SJC-PS-009. Banking Fraud Detection using Multimodal Transactions

Banking & Financial Services – Focused on fraud prevention using AI and multimodal data analysis.
Modalities:
Text: Transaction logs, account activity, and customer details.
Audio: Voice recordings from customer service calls or verification calls.
Image: ID documents, scanned checks, and profile pictures for identity verification.
Detailed Steps to Implement:
1. Data Collection & Integration:
 Gather transaction logs from banking systems (amount, location, device info, time).
 Collect voice call recordings from customer support or verification processes.
 Acquire ID images (government-issued IDs, selfies for KYC) for authentication checks.
2. Preprocessing& Feature Extraction:
 Text: Use NLP to detect anomalies in transaction patterns (e.g., sudden large transfers, unusual
geolocation).
 Audio: Apply speech-to-text and voice biometrics to verify identity and detect stress or
suspicious behavior.
Page 8 of 13
  Image: Use computer vision for ID validation (e.g., forgery detection, face matching with selfie).
3. Multimodal Fusion & AI Model:
 Combine features from text, audio, and image into a unified representation using multimodal
deep learning.
 Train fraud detection models using supervised learning and anomaly detection techniques.
4. Generative AI for Synthetic Fraud Scenarios:
 Use GANs or diffusion models to create synthetic fraud cases for model training and stress
testing.
 Simulate scenarios like fake IDs, voice spoofing, and transaction laundering patterns.
[Link] Detection Dashboard:
 Build an interactive dashboard for fraud analysts:
 Real-time alerts for suspicious transactions.
 Multimodal evidence (transaction details, voice snippet, ID image).
 Synthetic fraud scenarios for proactive risk assessment.
6. Reporting & Compliance:
 Generate automated fraud reports with confidence scores and recommended actions.
 Ensure compliance with banking regulations (KYC, AML, GDPR).
Expected Outcomes:
A fraud detection dashboard that:

 Displays real-time alerts with multimodal evidence.

 Provides synthetic fraud scenarios for training and preparedness.
 Generates detailed reports for compliance and investigation.

Page 9 of 13
Domain-4
Internet of Things (IoT)

NH-SJC-PS-010. Problem Statement: AI-Powered IoT Coaching System for Sports

Training
Background
Quality coaching is the difference between good and great athletes. However, personal coaches are
expensive, not always available 24/7, and inaccessible to millions of aspiring athletes. What if
technology could provide real-time, intelligent coaching that adapts to each athlete's needs, corrects their
form, designs personalized training plans, and motivates them like a human coach would?
The Challenge
Create an intelligent IoT coaching system that acts as a virtual sports coach - not just tracking metrics,
but actively training athletes through real-time guidance, technique correction, personalized workout
planning, and adaptive feedback based on performance and progress.
The Core Problem
Athletes Need a Coach, Not Just Data
• Passive devices tell you what happened, but don't help you improve
• Human coaches are expensive and not always available
• Generic training apps don't adapt to individual weaknesses or provide real-time correction
• Athletes train alone without immediate feedback on form, pacing, or technique

What a Real Coach Does (Your Device Should Too)

1. Observes and corrects technique in real-time
2. Designs personalized training plans based on goals and current level
3. Provides motivation and encouragement during tough workouts
4. Adapts training based on fatigue, performance, and progress
5. Teaches fundamentals and advanced techniques progressively
6. Prevents injuries by catching poor form or overtraining
7. Gives tactical advice during practice sessions

Page 10 of 13
NH-SJC-PS-011. Smart Home Intrusion Detection using Multimodal Sensors
IoT (Internet of Things) – Leveraging connected smart devices for home security.
Modalities:
Audio: Sounds captured by smart microphones (e.g., breaking glass, forced entry noises).
Image: Video frames or snapshots from smart cameras for motion and object detection.
Text: Metadata from sensors (timestamps, device logs, door/window status).

Detailed Steps to Implement:

1. Sensor Integration & Data Acquisition:
 Connect smart cameras, microphones, and door/window sensors to a central IoT hub.
 Stream audio and video data in real-time; collect sensor logs for contextual information.
2. Preprocessing& Feature Extraction:
 Apply noise reduction and audio feature extraction (MFCCs) for sound classification.
 Use image processing and deep learning (CNNs) for detecting human presence or unusual
activity.
 Combine sensor logs for event correlation (e.g., door opened + motion detected).
3. Multimodal Fusion & AI Model:
 Implement a fusion model that combines audio, image, and text features for robust intrusion
detection.
 Use anomaly detection algorithms or supervised learning for classification of intrusion vs.
normal activity.
4. Alerting & Reporting System:
 Trigger real-time alerts via mobile app or SMS when intrusion is detected.
 Generate incident reports with timestamp, detected modality, and confidence score.
 Attach visual evidence (snapshot or short video clip) for verification.
5. Dashboard & User Interface:

 Build a dashboard to monitor live feeds, intrusion alerts, and historical reports.
 Include analytics like intrusion frequency, time-of-day patterns, and device health status.
Expected Outcomes:
 Real-time intrusion alerts delivered to the homeowner’s device.
 Incident reports with detailed logs and visual/audio evidence.
 A user-friendly dashboard for monitoring and managing home security.

Page 11 of 13
NH-SJC-PS-012. Defect Detection in Manufacturing using Multimodal Sensors
Manufacturing – Leveraging AI and IoT for predictive maintenance and quality assurance.
Modalities:
 Image: Visual inspection of machine parts and products using cameras.
 Audio: Sound patterns from machines (e.g., grinding, vibration noise) for anomaly detection.
 Text: Machine logs, operational parameters, and maintenance history.
Detailed Steps to Implement:
1. Data Acquisition & Integration:
 Collect image data from cameras installed on production lines for visual defect detection.
 Capture audio signals from microphones placed near machines to monitor abnormal sounds.
 Gather textual logs from sensors and PLCs (Programmable Logic Controllers) for operational
context.
2. Preprocessing& Feature Extraction:

 Image: Apply computer vision techniques (CNNs) to detect cracks, misalignments, or surface
defects.
 Audio: Use signal processing (MFCCs, spectrograms) to identify unusual frequency patterns
indicating wear or malfunction.
 Text: Parse logs for error codes, temperature spikes, or vibration anomalies using NLP.
3. Multimodal Fusion & AI Model:

 Combine image, audio, and text features into a unified representation using multimodal deep
learning architectures.
 Train classification models to categorize defects (e.g., mechanical, electrical, surface).
4. Defect Classification & Alerting:

 Implement real-time detection and classification of defects.

 Trigger alerts for critical anomalies and schedule predictive maintenance tasks.
5. Report Generation & Dashboard:

 Generate automated maintenance reports with defect type, severity, and recommended actions.
 Include visual evidence (images) and audio clips for verification.
 Build a dashboard for supervisors to monitor machine health and defect trends.
Expected Outcomes:
1. A defect classification system that:

 Detects and classifies defects using multimodal data.

 Provides visual and audio evidence for each detected issue.
 Generates maintenance reports with actionable insights.
2. Additional features:
 Predictive analytics for future failures.
 Historical defect trend visualization.

Page 12 of 13
Domain-5
Innovation using Web 3.0

NH-SJC-PS-013. Front-Running Attack Detection & Mitigation System

Problem:
MEV (Maximal Extractable Value) bots constantly front-run regular users by inserting
transactions with higher gas fees, exploiting DeFi traders and NFT buyers. This leads to slippage,
unfair trades, and financial losses for everyday users.
Build a middleware DApp or dashboard that:
1. Detects suspicious mempoolbehavior, including sandwich attacks, front-running, and other
MEV patterns.
2. Shows flagged transactions, affected users, and potential slippage losses in real-time.
3. Optionally integrates a private transaction relayer (e.g., Flashbots) to allow users to submit
transactions safely.
4. Bonus: Includes a simulation mode where users can test transactions before submitting to
predict front-running risks.
Evaluation Criteria:

Metric Description Evaluation Criteria

Detection How well the system identifies % of correctly flagged malicious

front-running / sandwich attacks transactions vs. total simulated
Accuracy
in the mempool attacks

User Impact Clarity of information about Quality of dashboard visualization,

affected users and potential ease of understanding, actionable
Awareness
losses insights

Real-Time Speed at which transactions are Time from transaction broadcast

analyzed and flagged → detection (< few seconds ideal)
Performance

Page 13 of 13