Aastha Joshi

Identify ML-exploitable statistical vulnerabilities in the Speck lightweight cipher used in IoT and embedded defence communication systems.
• Applied classification models with entropy-based feature extraction; used SHAP explainability to isolate critical bit-level distinguishers in Speck cipher
output patterns.
• Designed differential distribution analysis pipeline distinguishing Speck variants from random permutations with statistically significant accuracy gains.

Developed PPO-based news recommendation systems improving user engagement by ~30%.
Built hybrid RAG and Thompson Sampling recommendation engine increasing product suggestion accuracy by 22%.
Deployed context-aware RAG chatbot using OpenAI Agent Kit for document-grounded query resolution.

Build scalable ML personalization and retrieval systems to improve recommendation quality and conversational AI accuracy over traditional baselines.
• Engineered PPO-driven news recommendation pipeline using RL to optimise content ranking dynamically; improved user engagement ~30% over
collaborative filtering baseline.
• Built hybrid RAG + Thompson Sampling product recommender balancing exploration-exploitation tradeoff; improved product suggestion accuracy by
22%.
• Deployed context-aware RAG chatbot with OpenAI Agent Kit for multi-step document-grounded query resolution at sub-second retrieval latency.

• Led 10+ member team organising 15+ technical workshops and hackathons; achieved 200% YoY engagement growth across college-wide tech events.
• Rotated across partnerships, marketing, and community engagement; built data-driven dashboards and strategic reports for the leadership team. received structured technical interview preparation, strategy, and industry navigation.

CNNs miss relational topology in medical images; GNNs can model spatial inter-tissue dependencies in graph-structured data for higher diagnostic accuracy.
• Implemented GCN models in PyTorch for medical image graph analysis; achieved 97% classification accuracy on benchmark healthcare imaging
datasets.
• Conducted ablation studies under Dr. Mallikharjuna Rao K, validating node-feature contributions on graph-structured clinical data classification.

Manual review of 10, 000+ daily transactions created audit bottlenecks and missed real-time data integrity failures in the auditing interface.
• Built Java/Spring Boot real-time audit dashboard monitoring 10, 000+ daily transactions, reducing manual review overhead by ~40%.
• Executed Postman API testing across 25+ test cases (95% UAT pass rate); surfaced 3 critical data integrity bugs in the production audit flow.