Agentic AI | LLM and agent evaluation | Reasoning, tool-use, multi-turn | AWS reInvent 2025

A framework for evaluating complex AI agents in production, with real-world examples from Amazon shopping, seller support, and advertising agents.. Elaborates on how to measure AI performance beyond traditional metrics, approaches for assessing language model reasoning, tool usage, and memory management. 

LinkedIn: Post 

Text2Image | Text2Video | Image2Image | Image2Video| Text2Speech

An all-in-one repository for open-source state-of-the-art Multimodal Large Language Models (MMLLM).

Github: https://github.com/yfgit2012/Multimodal-LLM 

ReAct agent | Multimodality | VLM fine-tuning | MCP | Strands Agents 

A next-generation multimodal agentic AI solution designed to automate and accelerate structural inspection processes, combining reasoning LLM, RL/GRPO fine-tuned VLMs, and MCP servers for seamless orchestration. The framework is developed with Strands Agents and Bedrock AgentCore, generates significant efficiency gains including 50% reduction in manual inspection efforts.

Blogpost: to be published

LLM post-training | Causal graph reasoning | RL | SFT and DPO

A multi-agent framework using reinforcement fine-tuning (RFT) to enhance the reasoning accuracy on expert-curated causal graphs, allowing smaller specialized models to outperform larger foundation models on domain-specific tasks, improving F1 score from 0.69 to 0.97. 

Paper: https://openreview.net/pdf?id=rbQLTUDZiH (accepted by AAAI 2026)

Graph RAG | LLM explanability| LLM knowledge base

A novel framework that combines RDF graph databases and RAG pipeline with LLMs to process natural language queries for precise audience targeting, while providing transparent reasoning through a planning-orchestration architecture. 

Paper: https://www.amazon.science/publications/an-explainable-natural-language-framework-for-identifying-and-notifying-target-audiences-in-enterprise-communication 

https://arxiv.org/abs/2508.05267

LLM migration | Data-aware prompt optimization | DSPy:MIPROv2

An LLM migration paradigm and architecture, including a continuous process of model evaluation and data-aware prompt optimization by MIPROv2,  iteratively optimizing LLM prompts using user-provided dataset and objective metrics. 

Blogpost: https://aws.amazon.com/blogs/machine-learning/improve-amazon-nova-migration-performance-with-data-aware-prompt-optimization/ 

LLM evaluation and benchmarking | Cost-performance analysis

An LLM evaluation framework, focusing on evaluating LLM performance, responsibility, infrastructure, and cost, driven by LLM tasks, providing Insights on LLM cost-performance for model selection and optimization. 

Github: https://github.com/aws-samples/sample-llm-task-ben-mig-aws 

RAG Optimization | SFT | RLHF/RLAIF | PPO,DPO | Compound AI System | DSPy

A continuous self-instruct fine-tuning framework and its pipeline implemented by DSPy. The framework generates a synthetic dataset from the domain knowledge base, and drives SFT, HITL, and RL with human alignment.

Blogpost: https://aws.amazon.com/blogs/machine-learning/llm-continuous-self-instruct-fine-tuning-framework-powered-by-a-compound-ai-system-on-amazon-sagemaker/ 

Github: https://github.com/aws-samples/amlc-2024-tutorial-continuous-fine-tuning-compound-ai 

Presented at Amazon Machine Learning Conference 2024

Blogpost:https://aws.amazon.com/blogs/machine-learning/how-amazon-finance-automation-built-a-generative-ai-qa-chat-assistant-using-amazon-bedrock/   

Machine Learning | AutoML | Explainable AI | VIX | Predictability | Forecasting | Quantitative Trading | Big Data | S&P 500 | Futures | US markets

Paper: https://www.tandfonline.com/doi/full/10.1080/14697688.2024.2439458 

Blogpost: https://aws.amazon.com/blogs/machine-learning/how-twitch-used-agentic-workflow-with-rag-on-amazon-bedrock-to-supercharge-ad-sales/ 

RAG | Instruction Tuning | RLHF/RLAIF

Presentation of RAG and LLM fine-tuning techniques, their advantages, limitations, and best-practice adoption strategies for various LLM tasks. Demonstration of advanced methods for optimizing RAG and fine-tuned LLM architectures for domain-specific applications. 

Proceeding: https://doi.org/10.1145/3637528.3671445 

Github: https://github.com/aws-samples/kdd-2024-domain-driven-llm-development 

LLM | HITL | RLHF/RLAIF | PEFT | LoRA | RAG  | AWS | SageMaker

Using Reinforcement Learning from AI Feedback (RLAIF) to scale up human-in-the-loop feedback data by LLM-as-a-judge, fine-tuned the 7B LLM via RL/PPO, improved overall accuracy to 89%, and reduced human annotation effort by 80%.

Blogpost: https://aws.amazon.com/blogs/machine-learning/improve-llm-performance-with-human-and-ai-feedback-on-amazon-sagemaker-for-amazon-engineering/ 

RAG | LLM | Instruction Tuning | PEFT | LoRA | AWS | SageMaker

A RAG optimization solution with SFT fine-tuned LLMs, improved accuracy from 50% to 81%. Launched enterprise-grade GenAI Q&A bot, reduced cost by 60% across mission-critical workflows.

Blogpost: https://aws.amazon.com/blogs/machine-learning/a-generative-ai-powered-solution-on-amazon-sagemaker-to-help-amazon-eu-design-and-construction/

Github: https://github.com/aws-samples/aws-rag-llmft-sagemaker 

An automated workflow using  AutoGluon AutoML framework on Amazon SageMaker, that performs high-accuracy (91%) theme detection that surfaces top customer contact reasons and enables faster, more effective issue resolution. 

Blogpost: https://aws.amazon.com/blogs/machine-learning/ai-ml-driven-actionable-insights-and-themes-for-amazon-third-party-sellers-using-aws/

An Auto Machine Translation and Synchronization (AMTS) system, an extensible, language-agnostic framework with customizable sub-pipelines that incorporate language characteristics and translator preferences, reduced human translation workload by 80%. 

Blogpost: https://www.amazon.science/blog/auto-machine-translation-and-synchronization-for-dive-into-deep-learning 

Research report and patent: https://inform.tmforum.org/research-and-analysis/case-studies/taiwan-mobile-transforms-its-oss-with-an-ai-ecosystem 

Patent US20170331673A1: https://patents.google.com/patent/US20170331673A1/ko  

Paper: https://www.sciencedirect.com/science/article/abs/pii/S0952197605000308 

Paper: https://ieeexplore.ieee.org/document/1215393 

Paper: https://ieeexplore.ieee.org/document/1005583