- aws
- ml
- ai
- cloud
- autonomous-driving
- llm
- agentic-ai
- rag
- multi-tenant
•
•
•
•
•
•
•
•
-
Invited Speaker at the KDD 2026 Day on AI Reasoning
Invited to speak on harness-guided reasoning for coding agents at the KDD Day on AI Reasoning, held August 12, 2026 in Jeju Island, South Korea, alongside speakers from UCLA, Virginia Tech, CMU, and AWS AI.
-
NKI-Agent: Domain-Specific Fine-Tuning and Agentic Tool Use for Neuron Kernel Generation
The first system combining domain-specific supervised fine-tuning with a compile-verify-fix agent loop for AWS Trainium kernel generation via the Neuron Kernel Interface (NKI). Tools take Claude Opus 4.8 from 6% to 77.3% pass rate, and a 3B-active SFT model reaches 25% at 1/100th the cost.
-
Five Rules for Multi-Agent Coding Teams — Derived From 27 Controlled Experiments
27 controlled experiments across 13 configurations reveal 5 operating rules for multi-agent LLM coding teams: smaller teams win, shared directory with scoped writes, nightly tests with failure injection, dedicated DevOps agent, N≥2 runs per config.
-
Guidance for Multi-Tenant Knowledge Base Management for Scalable RAG Applications on AWS
A centralized synchronization system that automatically distributes knowledge base updates across multi-tenant RAG applications — reducing operational overhead by up to 60% while ensuring tenant isolation and real-time content consistency.
-
FlashAttention on Trainium: Can an LLM Write Expert-Level Hardware Kernels?
We benchmark 10 NKI attention kernels on AWS Trainium, then show that an LLM agent can match a strong hand-optimized NKI attention kernel through iterative compile–verify–benchmark feedback — no RL, no fine-tuning.
-
Video-Text Temporal Localization via Multi-Scale Convolution and Dynamic Routing
A lightweight framework for video-text temporal localization that combines multi-scale temporal convolution and capsule-based dynamic routing to achieve accurate, efficient, and interpretable alignment between video segments and natural language queries.
-
AWS Guidance for AI-Driven Robotic Simulation and Training on AWS
Build an AI-powered robot training and fleet management system using Amazon Bedrock foundation models and AWS IoT. Combines imitation learning with NVIDIA Isaac on Amazon EC2 and reinforcement learning with edge-optimized reward functions to train robots for precise tasks and manage fleets at scale.
-
How BMW Group and Qualcomm built an automated driving platform on AWS
End-to-end automated driving platform combining Qualcomm's in-vehicle compute with AWS cloud services. Enables scalable data processing, large-scale simulation, and continuous L2+ feature development — from data collection in the vehicle to model training and validation in the cloud.