CV

MSc in Applied Computing (MScAC)

University of Toronto

Department of Computer Science

• Coursework (ongoing): Natural Language Computing, Reinforcement Learning, Numerical Optimization, HCI.

Bachelor of Engineering (Hons.)

Panjab University

Department of Computer Science & Engineering

• Coursework: Artificial Intelligence, Probability & Statistics, Data Structures & Algorithms, Database Systems.

Machine Learning Engineer I

Recursive Softpro

• AI-Powered Drone Survey Analytics & Conversational GIS Chatbot (PostGIS, LangChain, FastMCP)
  • Built an automated mapping pipeline converting raw drone imagery into high-precision 2D/3D maps (±1–3 cm accuracy), reducing processing time from days to minutes.
  • Deployed computer vision models to automatically detect road damage (cracks, potholes) and calculate surface areas (m²) for real-time repair prioritization.
  • Developed an interactive GIS chatbot enabling non-technical teams to query highway data, compare flight inspection cycles, and generate audit reports via simple chat.
• Agentic AI Orchestrator with RAG and Dynamic Tooling (Django, OpenAI, Pinecone)
  • Engineered a Swarm-driven multi-agent router with OpenAI function calling, dynamic tool resolution (registry + user-defined tools), and safe agent hand-offs.
  • Implemented end-to-end RAG using Pinecone (knowledge-base selection, top-k retrieval) and OpenAI chat completions with configurable parameters, tone, and language.
  • Developed production backend services with Django ORM and DRF for chat session persistence, tool registry management, and real-time system health monitoring.
• WindCraft: Resource-Constrained Scheduling Engine (Google OR-Tools, FastAPI, MongoDB, Docker)
  • Developed an automated construction scheduling engine using Google OR-Tools (CP-SAT solver), modeling resource capacity, seasonal productivity, and monthly commissioning targets to replace manual planning.
  • Built a self-healing scheduling system that detects solver infeasibility, diagnoses root causes (resource shortages, timeline conflicts), and automatically executes timeline extensions and crane assignment heuristics.
  • Architected a REST API microservice using FastAPI, MongoDB, and Docker to dynamically generate schedules for 5 to 50+ turbine projects, persist execution runs, and export structured Excel reports.

Language Research Analyst (Advisor: Prof. Pawan Goyal & Prof. Somak Aditya)

Indian Institute of Technology–Kharagpur

• Developed a data generation pipeline integrating manual annotation and LLM-assisted question synthesis (Google Gemini 1.5 Flash), creating over 4,300 high-quality Q&A pairs across lecture domains.
• Benchmarked state-of-the-art multimodal LLMs (Video-LLaMA, MPlug-Owl, etc.) under zero-shot, few-shot, and fine-tuned configurations to evaluate reasoning depth and generalization.
• Performed both quantitative (BLEU, ROUGE, METEOR) and qualitative (student preference) evaluations to assess answer relevance and pedagogical alignment.
• Authored paper on educational multimodal video question answering, accepted for publication at EMNLP 2025.

Research Intern (Advisor: Prof. Min Xu)

Carnegie Mellon University (CMU)

• Integrated Momentum Contrast for Unsupervised Visual Representation Learning (MoCo) with Video Vision Transformer (ViVIT) and ViT using weight inflation for 3D Cryo-ET classification.
• Proposed a novel pipeline for processing 3D data, resulting in improved performance and training efficiency for the classification task.
• Achieved an F1 score of 70.28 in open-set environments using ViVIT for classifying real Cryo-ET data and noise, setting a new benchmark in the field.
• Contributed to a research manuscript on self-supervised 3D Cryo-ET classification, involving architecture design, experimentation, and writing of related works.
• Written a book chapter on feature semantic segmentation, including membrane, template matching, and deep segmentation (In Review).

ACM Research Intern (Advisor: Prof. Sriparna Saha)

Indian Institute of Technology-Patna

• Collected a multimodal dataset (Hindi, Tamil, Bengali, Marathi) with paragraphs, summaries, and images from Large Scale Multi-Lingual Multi-Modal Summarization Dataset (M3LS).
• Extended the BART model with a visual-aware encoder to generate regionally contextualized summaries using both text and image data.
• Achieved a ROUGE-1 score of 0.266 on the Hindi dataset and 44.1% image precision using an image pointer for multi-output prediction.
• Work accepted at ICDAR 2024 based on multimodal summarization research with Indian regional languages.

Research Assistant (Advisor: Prof. Ram Sarkar)

Jadavpur University, Kolkata

• Compiled and preprocessed a Mendeley dataset with three classes (Covid-19, Pneumonia, Normal), applying data augmentation techniques to address overfitting.
• Extracted image features using Vision Transformer (ViT) and ResNet50 models.
• Developed an ensemble model by combining features from both ViT and ResNet50, and used XGBoost for classification, achieving 96.19% accuracy.
• Contributed to writing the results section of the paper, which was accepted at CPAMCS-2023.

Research Intern (Advisor: Prof. Sriparna Saha)

Indian Institute of Technology-Patna

• Collected a multimodal dataset of product images, customer reviews, and aspects from Amazon website.
• Extracted textual embeddings using BERT and visual representations using pre-trained CNN models such as VGG16 and ResNet.
• Developed a multimodal interaction model to learn the relationship between text, image, and product aspects.
• Achieved accurate complaint eligibility classification, with the work published at the European Conference on Information Retrieval (ECIR) 2023.

Research Intern (Advisor: Prof. Luo Wei)

National University of Singapore

• Utilized the U.S. Coast Vessel Traffic System dataset for Automatic Identification System (AIS) data analysis.
• Preprocessed and cleaned key navigational parameters including Latitude, Longitude, Speed Over Ground (SOG), Course Over Ground (COG), and Maritime Mobile Service Identity (MMSI).
• Applied clustering algorithms such as DBSCAN and K-Means to segment vessel trajectories and remove outliers.
• Developed and evaluated sequence-to-sequence models (RNN, LSTM, Bi-LSTM, GRU, Bi-GRU) for predicting vessel paths and assessing collision risks.

Research Intern (Advisor: Prof. Rajesh Kumar)

Indian Institute of Technology-BHU

• Designed and optimized concrete mix compositions by varying sand fines, cement, and plasticizer ratios to enhance strength, density, and durability.
• Implemented and compared multiple machine learning regression models — Multiple Linear Regression, Support Vector Regression (SVR), Decision Tree Regression (DTR), and Artificial Neural Network (ANN) — to predict compressive strength.
• Achieved the highest accuracy for both 7-day and 28-day strength predictions using the Decision Tree Regression model.

Multi-Agent LLM System for Automated App Development

Built a planner -> architect -> coder LangGraph pipeline turning plain-language specs into runnable project scaffolds with controlled recursion. Integrated Groq’s gpt-oss-120b through LangChain with Pydantic outputs and hardened sandboxed file/command access tool layers.

AI-Powered Knowledge Graph Builder

Built a knowledge-graph backend combining FastAPI with Ollama/OpenAI LLM pipelines for automatic entity and relationship extraction. Engineered production-grade ingestion (MIME validation, path traversal defense) and optimized SQLAlchemy persistence with Alembic migrations.

Intrusion Detection System (Panjab University)

Developed a deep learning-based IDS using NSL-KDD dataset, optimizing feature selection with RFE and Decision Tree classifier. Achieved 95% accuracy and 0.94 F1-score with CNN-LSTM model (accepted for publication in Elsevier Computers & Security).

E-Commerce Product Recommendation System

Leveraged pretrained ResNet50 backbone with Global Max Pooling to encode images into dense feature embeddings. Applied L2 normalization and brute-force k-Nearest Neighbors search to rank visually similar catalog items in a real-time Streamlit interface.