Selected work across privacy-preserving & federated ML, applied machine learning, and engineering systems. Each card links to its repository on GitHub.
Differentially private second-order federated optimization using a regularized Fisher Information Matrix, with a federated privacy accountant (Rényi-DP, subsampling amplification) and dimension-aware gradient clipping. Official implementation of the TMLR submission.
Federated domain-adaptation framework tackling domain shift and limited target data via functional-distance-based aggregation, aligning source/target models through mean gradient fields.
VAE-GAN decoder reconstructing visual stimuli from fMRI BOLD signals; FSL preprocessing (motion correction, GLM), ridge-regression voxel encoding and PCA/ICA dimensionality reduction.
Policy-gradient actor-critic over continuous parameterized action spaces, with the maximum-entropy principle and deterministic annealing for exploration control. B.Tech thesis.
Risk-scoring system using autoencoder residuals and reciprocity analysis with a hybrid NLP pipeline — BERT embeddings, TF-IDF and sequencing — to flag resume inconsistencies. Eightfold.AI Hackathon.
Identifying cryptographic algorithms from ciphertext using ML — feature extraction via NIST randomness tests and ensemble learning (Random Forest, Logistic Regression).
Natural language → structured JSON → Fusion 360 Python via RAG + LLM. SQL-backed ChromaDB retrieval with Qwen 2.5 Coder (7B) on Ollama. Built during the Philips R&I internship.
Native slicing and multi-axis G-code generation for an in-house 5-axis Ender (FANUC format) — built during the SONY SSUP internship for India's first 5-axis FDM printer.
Formula-Student vehicle simulation — suspension A-arm force analysis and bolt safety-factor computation across dynamic events (acceleration, skidpad, braking).