Projects

Activation-level detection of sandbagging, deception, and situational awareness in LLMs. Linear probes achieve 90-96% accuracy across Mistral, Gemma, and Qwen models. Includes steering vectors for runtime behavior control.

Adaptive Real-Time Enterprise Multi-agent Intelligence System. A production-ready framework that orchestrates structured debates between AI agents with hierarchical argument generation, jury-based evaluation, and integrated safety monitoring.

Runtime control of LLM agent behaviors through activation steering vectors. Modifying model outputs at inference time without retraining, demonstrating more calibrated control than traditional prompting.

Multi-agent trust and fraud engine that routes each decision to the cheapest sufficient layer. A five-level cascade escalates from deterministic rules through statistical ML, single agents and multi-agent debate to adversarial courtroom reasoning, with self-learning rule extraction and evolutionary adversarial testing hardening the defence over time.

Reference implementation of LLM inference acceleration techniques. Speculative decoding, tree speculation, EAGLE, Medusa, KV-cache compression, and diffusion efficiency optimizations.

High-performance GPU kernels for LLM inference operations using OpenAI Triton. Educational implementations demonstrating memory-bandwidth optimization techniques for transformer operations on A100 GPUs.

Distributed LLM evaluation framework built on Apache Spark for enterprise-scale model assessment. Handles millions of examples with statistical rigor, integrating seamlessly with Databricks infrastructure.

Framework for automated synthesis and verification of distributed systems. Bridges the gap between high-level specifications and provably correct implementations using compositional verification, CEGIS-based synthesis, and constrained RL.

Modern Java 21 implementation of the SMPP protocol. Built on Netty with virtual threads for high-performance SMS messaging at scale. The actively maintained replacement for Cloudhopper.

Production-ready bridge between SMPP protocol and Apache Kafka. Receives SMS messages via SMPP and publishes to Kafka topics. HTTP/2 REST API aligned with 3GPP TS 29.540 SMSF standards.

High-performance financial message processing tool for ISO 8583 payment protocol. Features CLI, Python SDK, and AI-powered message explanation. Supports all major card networks.

Privacy-preserving consent management protocol for LLM training data. Enables transparent opt-in/opt-out mechanisms with cryptographic verification for the GenAI era.

Open-source blockchain-based protocol for transparent personal data consent management. GDPR-compliant with automated compliance reporting and real-time privacy breach alerts.

Privacy-aware open protocol for non-repudiable location verification. Cryptographically secure location attestation with fully decentralized, tamper-resistant architecture.

Location verification framed as proof of presence rather than proof of identity. Uses zero-knowledge proofs so a party can verify where someone was without learning who they are. Pre-launch, with early access open.

Interactive visualizer for Manifold-Constrained Hyper-Connections (MHC), a novel architectural primitive for deep neural networks. Explore how MHC dials dynamically control information flow across transformer layers.

Canonical JSON Schema documents for the artifact formats used by DataDAG. Every schema is generated from the Pydantic model that defines it, so the published spec and the implementation cannot disagree. Additive changes keep their identifier; breaking changes take a new version.

Compiles REST (OpenAPI 3.x, Swagger 2.0) and SOAP (WSDL 1.1) service specifications into governance-aware MCP tool surfaces. Every field carries provenance, destructive tools are emitted disabled until a human approves them, and refusals are compiled in rather than configured. Evaluation is pre-registered and judged by deterministic state oracles.

Interactive reproduction of a theoretical physics treatise on symmetry conservation laws, originally presented at Kshitij 2006, IIT Kharagpur.