Coding Club

IIT Guwahati

EVEN
Semester
PROJECTS’25
 UNSUPERVISED TRAVELLING SALESMAN
PROBLEM

Problem Statement :
Replicate and extend the UTSP framework, which uses unsupervised learning and GNNs to solve the
Travelling Salesman Problem (TSP) without labelled data. The core idea is to train a GNN using a surrogate
loss to predict a heat map of edge probabilities, then apply local search heuristics to find a near-optimal
tour.

Goals:
Reimplement the UTSP model from scratch based on the paper and GitHub repository.
Modify the model using at least one of the following approaches:
Try a different GNN architecture than SAG.(e.g., GAT, GIN, GraphSAGE).
Use an alternative surrogate loss function.
Modify or enhance the local search algorithm.
→
Improve the T H transformation.
Scale to larger TSP instances efficiently.
Perform experimental comparison of the modified model with the original UTSP baseline on
multiple TSP sizes (n = 20, 50, 100, 200, 500).

Deliverables : Modified model version(s) with clear documentation of

Re-implemented UTSP model (baseline). changes.
Training and Evaluation Scripts for: Experimental Report (5–8 pages) including:
Heat map generation. Implementation details.
Local search. Challenges faced.
Metrics like tour length, time, and gap (%) Hyperparameters tried.
from optimal. Performance comparisons with charts/tables.
Code Repository (GitHub) with a good README Good README file:
file. Clear structure.
Instructions and usage guide.
TechStack/Frameworks :
Programming & Machine Learning : Python, with PyTorch or PyTorch Geometric for GNN-based
modeling and training.
Utilities & Visualization : NumPy, NetworkX, and Matplotlib for data handling, graph processing, and
visualizing TSP tours and edge heatmaps.

Shivansh Gupta Toshit Jain

Mentors 9560749686 6354642173
 PARK SMART

Problem Statement :
In urban environments, finding an authorized parking space near one’s destination is often a frustrating
and time-consuming task. Drivers typically spend several minutes circling blocks or parking lots,
contributing to traffic congestion, increased fuel consumption, and elevated stress levels. This inefficiency
not only affects individual commuters but also adds to broader urban mobility and environmental
challenges. There is a growing demand for a smarter, more efficient system that helps drivers locate,
reserve, and pay for parking spots with minimal friction.

Goals:
Build a mobile app that:
Identifies nearby authorized parking spaces using geolocation.
Displays real-time availability of parking slots.
Provides in-app navigation to guide users to the selected spot.
Optional Features:
Enables seamless digital payments to avoid physical transactions or paper tickets.

TechStack/Frameworks :
Frontend: Flutter / Compose / React native
Backend: Node.js / Firebase
Maps & Navigation: Google Maps API
Real-Time Updates: Firebase Realtime Database / WebSocket
Payments: Razorpay

Ayush Kumar Amit Yatika Jena

Mentors 9471495757 6201202256 9337854044
 MEDIGUIDE

Problem Statement :
Healthcare providers increasingly seek AI-assisted tools to deliver preliminary guidance and triage for
common medical inquiries. A conversational “e-doctor” can offer patients immediate, text-based responses
to symptom-related questions, while clarifying that it does not replace professional diagnosis

Transform an open-source, decoder-only Transformer model into a medically informed chatbot capable of
responding to patient questions with accurate, contextually appropriate, and professionally worded advice.
Compare three distinct fine-tuning strategies in terms of performance, resource utilization, and
deployment feasibility.

Goals:
Ingest a user’s free-text medical question and generate an answer that :
Adheres to recognized clinical guidelines.
Uses formal, professional language.
Includes disclaimers regarding non-substitutive informational use.
Supports fine-tuning modalities on identical training data like:
Prompt Tuning
Parameter-Efficient Fine-Tuning using the PEFT library (supports Adapter Training, LoRA, QLoRA and 8-
bit quantization)
Performance Report (PDF) including:
Dataset description and splits
Comparative results table summarizing ROUGE, PPL, latency, and model size.
Summary of trade-offs and recommended deployment strategy
Taking steps to ensure no patient-identifiable information is present; abide by HIPAA-equivalent
anonymization standards.

TechStack/Frameworks :
Machine Learning : Python, with PyTorch, HugginFace Accelerate
Transformers & Fine-Tuning : HugginFace Transformers, PEFT, BitsAndBytes

Shirshendu Sagnik
Mentors 9674397828 9874693499
 CVALIGN

Problem Statement :
In today’s competitive job market, companies receive hundreds of CVs for each role, yet hiring decisions
remain heavily reliant on manual screening. This process is inefficient, inconsistent, and prone to human
bias. To streamline candidate evaluation and improve hiring quality, develop a CV evaluator website that
uses a RAG (Retrieval-Augmented Generation) LLM model to assess CVs against specific job role
descriptions provided by companies.

The system should enable companies to upload job role descriptions—defining required skills, preferred
experiences, and traits—into a centralized platform. Job seekers’ CVs are then automatically evaluated that
assigns a relevance score and generates personalized feedback, such as strengths, weaknesses, and role-fit
explanations. The result is a structured, explainable report that helps recruiters quickly identify the best
candidates.

Goals:
A web form where companies enter role-specific requirements and preferred qualifications.
Role-Based Access Control
Different permissions for recruiters, hiring managers, and admins.
Extracts and structures content from uploaded CVs (PDF/DOCX).
Stores all the upload CVs in a cloud storage in a seamless manner
Uses a RAG model to:
Score each CV based on relevance to the job description.
Generate textual feedback explaining strengths and weaknesses.
A web dashboard showing:
Ranked candidates with scores.
Detailed feedback per CV.
Filtering and comparison features.

TechStack/Frameworks :
Machine Learning : Python, with PyTorch, Hugging Face Transformers, Langchain and Langgraph,
PIneCone or FAISS for vector database storage
Web : HTML/CSS + ReactJs, FastApi , any cloud storage(like Cloudinary)

Krish Shirshendu
Mentors 7488665219 9674397828
AXIS-ALIGNED ENCLOSURE OPTIMIZATION

Problem Statement :
Your nation has begun urban redevelopment in the futuristic mega-city of Axion Prime, a vast, grid-
structured metropolis built on top of ancient technology. The city is populated with various buildings
scattered across the land, each associated with a cost — a monetary burden for including that building
within a regulated zone.As the chief architect of Axion Prime's zoning council, your task is to design a
containment boundary — a simple, closed polygon — to enclose at least K buildings

Goals:
Total Cost = Perimeter of the polygon + Sum of weights of enclosed buildings
Polygon must be:
axis alligned (its edjes should be parallel to x and y axis).
vertices can be floating points.
A building is considered enclosed if it lies strictly inside or on the boundary of the polygon.
The polygon must be simple — meaning non-self-intersecting and closed.
Design an axis-aligned, simple polygon that encloses at least K buildings, and minimizes the total
cost as defined above.

Input Format: Constraints:

The first line contains two integers N and K — the 1 <= k <= N <= 10000
total number of buildings and the minimum 0 <= xᵢ ,yᵢ <= 10000
number that must be enclosed. wᵢ can be negative
The next N lines each contain three integers: xᵢ yᵢ wᵢ Polygon must be simple,closed and axis alligned
— the coordinates and cost weight of the i-th
building.

Output Format:
The first line must contain a floating-point number C — the minimum total cost, printed with at least six
digits after the decimal point
The next lines should describe the polygon, one edge per line.
Each line contains four real numbers: x1 y1 x2 y2, representing an edge of the polygon from point (x1, y1)
to (x2, y2)
Edges must be listed in order, such that consecutive edges form a connected path and the last edge
connects back to the first
Each edge must be either horizontal or vertical

Parth Aher Vaibhav Singh

Mentors 9527409158 6200449080
 CLEAR-VISION

Problem Statement :
Image degradation is common in real-world visual data due to noise, compression artifacts, occlusion, or
partial corruption. This project focuses on building an image restoration system that can recover and
reconstruct high-quality images from such degraded inputs using deep generative models.

The goal is to design a model that takes a corrupted image and restores it to a clean, high-fidelity version.
You are encouraged to explore and compare multiple generative modeling approaches including
Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), or Diffusion Models. A key
component of the task is to develop your own corruption algorithm to simulate realistic and challenging
image degradation scenarios.

Deliverables :
Selenium-based image web scraper
Dataset description and splits
Image corruption module
Implementation of any of the three restoration models:
VAE
GAN
Diffusion Model
Evaluation metrics computation including:
PSNR
SSIM
LPIPS
Inference latency
Web interface or mobile interface for upload and restoration preview

TechStack/Frameworks :
Machine Learning : PyTorch or TensorFlow
Web Interface : Streamlit (for Website), Flutter (for Mobile)
Web Scraping : Selenium

Sagnik Prakhar
Mentors 9874693499 9305415358
 SCHOLARSHIP FINDER

Problem Statement :
Build a Scholarship Finder tool that scrapes available scholarships from various educational portals,
government websites, and scholarship platforms. The tool will match scholarships to students based on
their personal profiles such as their course, grades, location, and interests. This will help students find
the most relevant opportunities automatically, making it easier for them to apply for scholarships.

Goals:
Student Profile Section- Input fields for course of study, GPA/grades, location, and optional criteria
like income status or special categories..
Scrape Scholarship Data: Develop a web scraper to automatically collect scholarship information
from :-
Major platforms (e.g., Chegg, Scholarships.com, Fastweb).
University-specific scholarship portals.
Niche platforms for field-specific or demographic-specific scholarship.
Categorize & Match Scholarships: Develop a web scraper to automatically collect scholarship
information from :-
Amount : (Full, Partial, or Small Grants)
Eligibility: Course, GPA, location, special criteria
Deadline: (with notifications for urgent deadlines)
Use a matching algorithm to align scholarships with individual student profiles.
Generate Personalized List: Display a list of scholarships tailored to the student, ranked by:
Relevance to the student’s profile.
Deadline urgency
Scholarship amount.
Each entry includes eligibility, deadline, amount, and a direct application link..

TechStack/Frameworks :
FrontEnd: HTML+CSS ,ReactJs,
Backend: Node.js+Express/FastApi
Scraping: BeautifulSoup / Scrapy
Sentiment Analysis: TextBlob / VADER
Database: MongoDB.

Aditya Samal Krish

Mentors 7815018451 7488665219
 COMPANYGROW

Problem Statement :
CompanyGrow is a smart workforce development platform that enables organizations to nurture
employee growth through tailored training, intelligent project allocation, and real-time performance
rewards. It matches employees to projects based on skills, tracks training progress, and rewards
achievements with badge-based bonuses . With live analytics and a skill-driven ecosystem, it keeps
teams engaged, productive, and continuously evolving.

Goals:
Training Course Catalog:
Admins/Managers create/manage courses.
Employees browse, filter & enroll based on interest and skill level.
Employee Profiles:
Profiles include skills, experience & training progress.
Visual progress bars for milestones & completed courses.
Skill-Based Project Allocation:
Admins/Managers assign projects based on skills/training
Visual progress bars for milestones & completed courses.
Stripe Payment Integration:
Employees earn badges/tokens on course or project completion
Badges are tracked in real-time and factored into performance-based bonuses
Performance Analytics
Track training/project metrics and badge counts.
Visualize progress through graphs showing monthly/time-based performance.
Exportable reports (CSV/PDF) for reviews and insights (optional)
Admin Dashboard:
Central panel to manage courses, employees, and projects.

TechStack/Frameworks :
Frontend: HTML+CSS, Reactjs
Backend: Node.js, Express.js, Cloudinary (for media storage)
Database: MongoDB or PostgreSQL
Payment: Stripe (badge-based reward system, real-time processing)
PDF Generation: pdf-lib or jsPDF (for generating reports)
Analytics & Graphs: Chart.js or Recharts (for performance visualizations)

Aditya Samal Krish

Mentors 7815018451 7488665219
 ZCODER

Problem Statement :
ZCoder is a collaborative coding platform that enhances learning through interaction and personalization.
Users can create profiles, bookmark coding problems with solutions, and engage with others through
comments and feedback. With features like a live code editor and interactive rooms, The Platform can be
used for practice and peer collaboration.

Goals:
Personalized Profiles – Allow users to build their unique coding identity and connect with peers
Collaborative Learning – Enable users to share solutions, give feedback, and learn from one
another.
Efficient Practice – Streamline coding practice with accessible tools and saved resources.

Deliverables :
User Profiles Section– Detailed profiles with saved problems, solutions, and contributions..
Code Bookmarking – Save and organize coding problems with tags of the topics of the problem
Interactive Rooms – Real-time chat spaces for peer discussion and collaboration.
Inbuilt Code Editor – Integrated editor with real-time output for hands-on coding practice.

TechStack/Frameworks :
Frontend: HTML+CSS, Reactjs / NextJs
Backend: Node.js, Express.js
Database: MongoDB
Real time communication: Web-Sockets

Aditya Samal Krish

Mentors 7815018451 7488665219
 AGRISCAN

Problem Statement :
AgriScan is a smart plant disease detection mobile application designed to support farmers by leveraging
mobile technology and machine learning. Plant diseases significantly impact crop yields and food security,
especially for smallholder farmers lacking access to expert diagnostics. Early and accurate detection is vital to
avoid misdiagnosis, reduce harmful pesticide use, and ensure effective treatment. AgriScan enables disease
identification through images captured on smartphones, using an on-device machine learning model for fast,
cost-effective diagnosis even in remote areas.

Goals:
Early Detection – Identify plant diseases quickly using smartphone images and machine learning.
Actionable Remedies – Provide users with treatment suggestions, both natural and chemical.
Farmer Empowerment – Educate users with care tips and enable community knowledge sharing.

Deliverables :
Disease Detection via Camera – Capture or upload images for instant disease diagnosis using a CNN
model.
Treatment Recommendations – Remedies and prevention strategies for identified diseases.
Scan History Tracker – Maintain a record of scanned images, predicted diseases, and treatments applied.
Optional Features – Multilingual support, farmer forums, cloud sync, and admin dashboard for agri
officers.
Use of augmentation and class balancing techniques
Model training: CNN, MobileNetV2, ResNet18, EfficientNet
Metrics: Accuracy, Precision, Recall, F1-Score, Confusion Matrix, Model Size, Inference Speed
Final Report (PDF): Dataset insights, model evaluations, deployment strategies

TechStack/Frameworks :
Frontend: Flutter (Android & iOS support)
Backend (Optional): Firebase (Authentication, Firestore, Realtime Database, Storage)
Machine Learning:
TensorFlow (model training & conversion to .tflite)
tflite (Flutter integration)
scikit-learn, Matplotlib (data processing & visualization)
Hugging Face Datasets (for dataset access)
Mapping (Optional): Google Maps API
Realtime Updates: Firebase Realtime Database

Yatika Jena Yash Ayush Kumar

Mentors 9337854044 9137515319 9471495757