Welcome to our newest installment of EDGE AI Workshops – providing you with hands on knowledge to build the latest edge AI technologies, taught directly from industry and academic experts – FOR FREE.
———————————————
WORKSHOP: Building a Privacy First Voice Kiosk: Local RAG & LLMs on Arduino UNO Q
START TIME: 8:15AM PST
Hosted by Dr. M. Fekri (Moorcheh) & A. Richetta (Qualcomm-Arduino)
NOTE: Arduino is providing 10 UNO Q boards to builders who attend and interact during the live webinar. Attendees will also receive complimentary access to Moorcheh Edge (up to 10,000 vectors). Together, it’s an ideal hardware and software setup to build Moorcheh information theoretic search engine and RAG on Arduino to build your local AI assistants.
AUDIENCE: Edge AI developers and innovators building offline, privacy preserving AI products without cloud dependencies.
GOALS:
(1) Deploy a fully local, hands free AI assistant managing concurrent workloads within a ~3.6 GB RAM limit.
(2) Use Moorcheh Edge 1 bit embedding compression for ~17 ms semantic search on device.
(3) Orchestrate the offline stack: local LLM (Llama 3.2 1B), Whisper (STT), and Piper (TTS).
(4) Transition to a standalone appliance running continuously on USB-C power without cloud APIs.
PREPARATION: None mandatory. To replicate: Arduino UNO Q (4GB), USB mic/speaker, host PC on the same Wi-Fi, and the provided walkthrough repo.
ACTION: Swap the demo catalog for your own data and run the complete open source pipeline with zero cloud fees.
TIME: 1 hour (webinar + live demo)
————————————————–
WORKSHOP: Training Free Food Understanding with Vision Language Models at the Edge
START TIME: 9:30AM PST
Hosted by D. Pau and A. Spagnolo (STMicroelectronics) and A. Nagendra (Qualcomm).
TARGET AUDIENCE: Edge AI developers, computer vision engineers, embedded developers, and technical decision makers interested in running multimodal AI pipelines locally on constrained hardware.
GOALS: By the end of the session, attendees will understand:
(1) How a training free food understanding pipeline combines image text retrieval and vision language models without task specific fine tuning.
(2) How image text retrieval reduces large ingredient and caption vocabularies to a small candidate set before invoking a more expensive VLM.
(3) How confidence based gates skip unnecessary model calls to reduce latency while preserving prediction quality.
(4) How this architecture supports three concrete tasks:
(4.1) Ingredient recognition – retrieve likely ingredients and refine the final selection with Qwen3-VL-4B-Instruct.
(4.2) Dish caption retrieval – retrieve candidate descriptions and rerank them with Qwen3-VL-Reranker-2B.
(4.3) Calorie estimation – estimate ingredient quantities or portion sizes and combine them with nutritional information.
(5) How the full pipeline was optimized and deployed locally on both NVIDIA Jetson AGX Orin and Qualcomm Dragonwing IQ-9075 EVK.
PREPARATION REQUIRED:
(1) Nothing mandatory to follow along.
(2) To reproduce the demo: an NVIDIA Jetson AGX Orin and a Qualcomm Dragonwing IQ-9075 EVK, plus the project repositories and setup instructions shared with attendees.
DURATION: 1 hour (technical presentation + live on device demo)
————————————————–
WORKSHOP: When Buildings Speak in Codes
Neural Codec & On Device Qwen3 for Vibration Diagnostics
Hosts: N. Kordjukovs & D. Pau (STMicroelectronics), A. Nagendra (Qualcomm)
START TIME: 10:45AM PST
AUDIENCE: Edge AI developers and engineers building private, sensor driven diagnostics without sending raw data off device.
GOALS:
(1) PIPELINE: Learn how the codes_v3 pipeline converts 10 second sensor windows into 250 single token symbols per sensor via a residual VQ codec.
(2) COMPARISON: Compare 1 reference STWIN.box (STEVAL-STWINBX1) against 5 targets using codec symbols and broadband levels.
(3) SCENARIOS: Explore live 6 board USB monitoring (running Qwen3-4B on the IQ-9075 NPU), immutable recorded replay, and operator triage.
(4) PRODUCTION: Keep raw data local, hash verify files, apply strict schema validation, and use fail closed error handling.
PREPARATION REQUIRED: None mandatory. Live replay requires: Qualcomm IQ-9075 EVK, 6 STEVAL-STWINBX1 boards, a USB hub, and the provided repo.