Research Papers

Trilochan sharma

A measurement-first look at how multi-agent LLM systems should share context. Across 33 real Claude generations, reference-passing (TOAP) and a competent summarizing orchestrator both cut tokens with zero accuracy loss — but summarization actually wins on raw token savings. TOAP earns its place anyway: it's lossless (8/8 vs. 6/8 when a downstream task needs a fact the summary dropped, replicated on two open non-Claude models) and it blocks 40/40 prompt-injection attempts via a capability lattice with zero false positives. A separate KV-cache transfer experiment across seven open models shows up to ~180x prefill speedup, gated by attention architecture.

trilochan sharma

ContextForge: A Research-Grade Nexus for Secure, Persistent AI Memory Developed a "Defense-in-Depth" architecture to solve the "Stateless RAG Gap" in AI Agents. By implementing a dual-signal security gate using Shannon Entropy and LZ Compression, ContextForge blocks adversarial injections and repetitive attacks while simultaneously reducing provider failover latency by 68.9%. Validated via the 375-test OMEGA-75 suite with an 80.7% Composite Safety & Performance Index