Aditya Yadav

• One of 2 CSE seats selected for a competitive 6-month paid internship on defence vehicle systems R&D.
• Working on an AI-based Vehicle Health Monitoring System (VHMS) for engine and gearbox health.
• Proposing a hybrid architecture: XGBoost/Random Forest on ECU sensor data, CNN on vibration/acoustic signals, LSTM/Transformer for time-series degradation prediction, and autoencoder-based anomaly detection.

• Built GRAVISENSE — a passive, slope-adaptive eddy current torque resistance attachment for mobility devices — as part of a multidisciplinary team.
• Owned the full pipeline: problem identification, CAD design, prototyping, fabrication, and testing.
• Won 3rd Prize (Best Project) among all projects presented during the internship.

• Built and evaluated a salary prediction model using CNN and GAN on the IBM Adult dataset (48K records).
• Achieved 87% classification accuracy — 11% above a logistic regression baseline.

• Developed a deepfake detection system using a CNN-based binary classifier on the FaceForensics++ dataset.
• Achieved 93% accuracy distinguishing real vs. AI-generated images.
• Reduced the false-negative rate by 18% vs. a baseline VGG16 through data augmentation and domain-specific fine-tuning.

Privacy-first mobile chat app built for a national security hackathon. Passed adversarial red-team evaluation with zero data-leak findings.

• End-to-end encryption via Libsodium, ephemeral TTL messaging
• Serverless P2P architecture — no central server
• Zero data-leak findings in adversarial red-team evaluation

Full-stack web app centralizing internships, research, and hackathons for students. 200+ users onboarded in first month.

• 200+ students onboarded within first month of launch
• Advanced search, filtering, bookmarking, and user dashboards
• Reduced opportunity-discovery time by ~60% vs manual search

Biomedical RAG pipeline using Bio-Mistral LLM with vector database retrieval for context-aware medical responses.

• Generated embeddings from 10,000+ biomedical documents, stored and queried via vector DB
• Achieved 0.82 ROUGE-L on a held-out QA set
• Built as part of a Guvi certification project (Jan 2025)