Research
Dipankar Sarkar's research, organised by area. Each area is a pillar page with the relevant papers, FAQ, and cross-links to the consulting practice and the engineering work.
Federated Learning
Fed-Focal Loss (93 citations), CatFedAvg, privacy-preserving distributed ML. The research area Dipankar is best known for.
GPU Kernel Correctness
The Correctness Illusion in LLM-generated GPU kernels. 26-op corpus with op-schema-aware seeded fuzzing.
DePIN
Decentralized Physical Infrastructure Networks. The Generalised DePIN Protocol and the theoretical framework.
Multi-Agent Systems
Multi-agent coordination in software engineering. The new inner loop: five agents converging on a single correct solution.
RAG & Knowledge Retrieval
Retrieval-Augmented Generation, knowledge conflicts, and the Input-Regime Audit framework for production RAG systems.
MEV Mitigation & Ethereum
FairFlow Protocol — equitable Maximal Extractable Value (MEV) mitigation in Ethereum.
Decentralized Deepfake Detection
A blockchain-based network for dynamic deepfake detection algorithm management.
Privacy-Preserving AI
Federated learning, differential privacy, secure aggregation, and the Cryptography + ML intersection.