Cut the cost of agentic LLM memory by deciding what to write before calling an LLM. SAGE is a closed-form novelty gate that routes new facts to ADD, redundant ones to SKIP, and only sends genuinely ambiguous cases to an LLM — 3.4× cheaper, 2.5× faster writes.
A novel teacher‑free distillation method that replaces live teacher inference with compact precomputed random projection teacher fingerprints, enabling storage-efficient, high-performance knowledge transfer and enhanced student training.
A curated research atlas for Efficient ML, AI Infrastructure, and Physical AI
Quantized Defect Correction for low-precision neural operators: precision-sensitive failure diagnosis, selective recovery, and A100 inference evaluation.
A hands-on lab comparing PTQ INT8, magnitude pruning, and knowledge distillation on small CNNs.
Heavy-tail-aware weight-only post-training quantization for vision models, with layer diagnostics, robust quantizers, and reproducible experiments.
Official Repo of OpenArchX Framework.
First 1-bit (BitNet b1.58) recursive reasoner for Sudoku-Extreme - distilled from a 7M-param FP TRM teacher into a 1.4 MB ternary student
Layer-adaptive quantization of diffusion models using Fisher Information (LAQFI), presented at IEIE 2024.
CPU-first, evidence-first static analysis for suspicious Python source: AST to MalIR to selective micro-model inference.
Paper, results & materials for our NeurIPS 2025 (Workshop on Efficient Reasoning) paper on cost-aware multi-agent orchestration — LoRA-Guided PPO.
Model compression engine built from first principles - magnitude pruning, per-layer sensitivity scanning, and linear quantization (int8/int4/fp8/fp4), benchmarked as a full prune x quantize matrix.
Pilot study on post-training quantisation effects for low-resource NMT (Hausa, Swahili, Yoruba) across three model families on consumer GPUs. Prepared for PhD applications 2027.
O(N) Linear Complexity Transformer - 125x faster than standard attention using cumulative sum trick. Pure PyTorch implementation.
Source for my GitHub profile README and its banner graphics: a PhD researcher in Computer Science and Engineering at the University of South Florida working on efficient and generative AI for computer vision and imaging, with links to research such as BAP-MOS, publications, tools and ways to get in touch.
KronLLM: a 10M-parameter bilingual English/Russian language model whose weights are Kronecker sums of rank-1 factors instead of dense matrices. Q10 (9,938,886 params) is the model; the 2-3M TinyStories runs are the development bench. Full research log, negative results included.
FP8 quantization and token eviction of the KV cache for Llama-3.2-3B (vLLM, kvpress)
Notes and Labs from MIT 6.5940, (Fall 2023) lecture.
To associate your repository with the efficient-ml topic, visit your repo's landing page and select "manage topics."