agenttrace
A non-invasive debugging framework for multi-agent LLM systems that reconstructs causal fault graphs from execution log…
Building the future, one commit at a time. Here are some of my open-source projects and research work.
A non-invasive debugging framework for multi-agent LLM systems that reconstructs causal fault graphs from execution log…
GNN + LLM hybrid recommendation with grokking detection using LightGCN/GraphSAGE and Qwen2.5-3B.
30+ layer contribution metrics from 7 categories unified in one toolkit, with bridges for Torch-Pruning and PEFT LoRA r…
Production-grade Grouped Query Attention (GQA) in Triton with composable attention patterns: 8 variants via tl.constexp…
CoNLL 2026 submission: Evidence-Based Dominance theorem proves evidence verification is theoretically optimal for elimi…
Targeting JASIST/ACL: The first citation verification system that goes beyond existence checking to full semantic claim…
ML-aware caching for recommendation systems: 45% cache hit rate, 85% latency reduction on hits, and under 3% NDCG degra…
KV cache optimization achieving 10-16x faster warm latency vs vLLM, validated across 84 tests on TinyLlama, Mistral-7B,…
Three-layer cognitive memory for LLM agents -- working, episodic, and semantic -- with 473 passing tests, hybrid retrie…