Projects, Research & Case Competitions

Contributions

• Addressed the scalability and localization challenges of celebrity-led advertising by developing a framework that adapts a single brand campaign for different retailers, locations, and promotional offers.
• Designed a brand-to-retailer personalization workflow enabling MSMEs to generate localized advertisement variants without requiring a separate video shoot for every variation.
• Architected a two-stage multimodal AI pipeline integrating 5+ components across speech recognition, voice generation, lip synchronization, facial restoration, and content moderation.
• Developed spoken-entity personalization capabilities to replace brand names, location references, and promotional offers while preserving the original speaker’s voice identity and synchronized lip movements.
• Enabled multilingual advertisement localization with support for 22 Indic languages, targeting regionally tailored advertising across diverse linguistic markets.
• Engineered an evidence-based benchmarking framework using 640 speakers across two language families to evaluate model suitability for multilingual speech applications.

Recognition

Colloquium 2026

Second Prize

Recognized with second prize at Colloquium 2026.

Key Highlights

• Modeled a multi-stage supply chain network spanning plants, mother hubs, Clearing and Forwarding Agents (CFAs), and customer demand.
• Developed inventory planning methodology using average demand, forecast variability, lead times, safety stock, reorder points, and days of cover.
• Formulated a Mixed-Integer Linear Programming (MILP) model to minimize modeled production, transportation, and unmet-demand penalty costs.
• Incorporated production batch-size and capacity constraints to represent fixed 25 kL production batches and plant- and category-level limitations.
• Evaluated a 70% Overall Equipment Effectiveness (OEE) scenario to assess its impact on production capacity, total modeled cost, and unmet demand.
• Identified structural capacity bottlenecks and translated optimization outputs into actionable production planning and resource allocation insights.

Recognition

Competition result

National Runner-Up

Castrol PowerUp 4.0 Supply Chain Track case competition.

Contributions

• Developed a real-time vehicle-collision detection system using YOLOv11 and CCTV surveillance footage to identify road accidents across 8+ accident categories.
• Optimized the detection workflow to achieve a reported processing cycle of 8 ms per frame, supporting low-latency video analysis.
• Built a FastAPI backend for accident-detection services and incident event handling.
• Implemented WebSocket-based real-time alerts to communicate detected incidents to the monitoring interface.
• Developed a Next.js and TypeScript frontend with a three-stage incident-verification workflow for reviewing and responding to detected accidents.
• Integrated authentication, PostgreSQL, and Prisma to support application access and persistent data management.

Reported model performance

mAP@50

99%

Reported object-detection performance across the evaluated accident classes.

mAP@50–95

89.3%

Reported mean average precision across stricter IoU thresholds.

Precision

96.8%

Reported proportion of predicted accident detections that were correct.

Recall

96.4%

Reported proportion of ground-truth accident instances detected.

Contributions

• Developed the AI/ML component of a larger automated question paper generation system designed to support assessment workflows at an autonomous institute.
• Curated and preprocessed a dataset of 16,200+ questions from multiple educational sources using data cleaning, tokenization, and augmentation.
• Fine-tuned RoBERTa to classify questions across six Bloom’s Taxonomy levels, comparing its performance against BERT and DistilBERT baselines.
• Achieved 93.7% classification accuracy through model optimization and hyperparameter tuning.
• Integrated the ONNX-optimized classification model with a FastAPI backend to support the wider question paper generation workflow.

Model performance & workflow impact

Classification Accuracy

93.7%

Achieved by the fine-tuned RoBERTa model across six Bloom’s Taxonomy levels.

Manual Classification Effort

95% reduction

Reported reduction through automation of taxonomy-driven question classification.

Key Highlights

• Built a one-command portfolio experience that launches using npx vishalrmahajan with Node.js 14+.
• Created colorful terminal styling and ASCII banners using chalk and cfonts.
• Used terminal-kit for interactive widgets and keyboard navigation.
• Integrated terminal-link for clickable terminal hyperlinks.
• Used open for cross-platform link opening.
• Provided zero-install execution through npx and a global-install option for users who want to run the CLI directly.