Comprehensive, scalable ML inference architecture using Amazon EKS, leveraging Graviton processors for cost-effective CPU-based inference and GPU instances for accelerated inference. Guidance provides a complete end-to-end platform for deploying LLMs with agentic AI capabilities, including RAG and MCP

Implementation of the model "Hedgehog" from the paper: "The Hedgehog & the Porcupine: Expressive Linear Attentions with Softmax Mimicry"

A fully offline NLP pipeline for extracting, chunking, embedding, querying, summarizing, and translating research documents using local LLMs. Inspired by the fictional mystery of Dr. X, the system supports multi-format files, local RAG-based Q&A, Arabic translation, and ROUGE-based summarization — all without cloud dependencies.

Variance-stable routing for 2-bit quantized MoE models. Features dynamic phase correction (Armen Guard), syntactic stabilization layer, and recursive residual quantization for efficient inference.

🛠️ Prototype for the Kaggle x Google Agents Intensive - Capstone Project Community Hackathon 2025. TaskFlowr is a 3-agent system built using the OpenAI ADK & Google AI that automates general business operations.

CodeSilver is an ambient AI prototype that translates clinical language into operational healthcare insights using synthetic data and public CMS rules, helping reduce documentation gaps without interrupting clinicians.

🩺 Translate clinical discussions into real-time hospital operations data to reduce errors and improve workflow efficiency.

Raw logs of Claude Code running on local Qwen3.5-27B (llama.cpp). Builds a Python todo app with 50 tests. Real-world performance data: 30 min, cache thrashing, 38 t/s generation.