← All events

Past eventOrganised by our community

Pie & AI: Bangkok - Embedded AI

A DeepLearning.AI community meetup hosted by AGICAFET

When
13:00–18:00 · Bangkok time (ICT, UTC+7)
Where
Cleverse Rama 9Bangkok
Hosted by
AGICAFET ↗ with DeepLearning.AI Pie & AI community

This event has already taken place. The slides stay online here.

Pie & AI title slide: a sweet meetup for deep learners, by DeepLearning.AI

Sessions & slides

Talk

NVIDIA Isaac & Ecosystem Future of the Physical AI

Natdhanai PraneenatthaveeAGICAFET

AGICAFET has joined the NVIDIA Inception Program. Natdhanai Praneenatthavee walks through how robot intelligence scales from specialists to generalists, why training data is the grand challenge of robotics, and NVIDIA’s end-to-end workflow — collect, curate, simulate, train and evaluate in Isaac Lab, deploy with Isaac ROS — then Isaac GR00T, agents that combine skills, tools and context for physical AI, and a robot-arm demo of how GapONet could fit into RL policy training.

In the deck

  • AGICAFET joins the NVIDIA Inception Program: making AI inference faster, leaner and cheaper
  • Scaling robot intelligence for the real world: from specialist to generalist, and the anatomy of autonomy
  • Training data is the grand challenge of robotics, and compute is data: teleoperation, simulation, internet and synthetic data, world foundation models
  • The end-to-end robotics workflow: data collection, curation and synthesis with Omniverse and Cosmos, training in Isaac Lab, evaluation in Isaac Lab Arena, deployment with Isaac ROS
  • NVIDIA Isaac GR00T: the end-to-end workflow for humanoid development, from data to deployment
  • Agents at work: skills, tools and context, and the agent tools and open models for physical AI
  • Demo: how GapONet could be used with RL policy training, on a small robot arm
Open PDF24 pages · 3.9 MB
Talk

Your GPUs are doing the same work twice: KV-cache reuse for faster, cheaper LLM inference

Dai TranTensormesh

On hardware you own, you pay to recompute the same context on every request. Dai Tran explains what the KV cache is, why prefix caching stops short, how LMCache, CacheGen and CacheBlend keep and reuse it across requests and instances, the measured gains (about 3× faster time-to-first-token at equal quality), and how the layer drops into a vLLM stack on-prem.

In the deck

  • The problem: on hardware you own, you pay to recompute the same context
  • The blind spot: everyone watches GPU utilisation, almost nobody watches KV-cache hit rate
  • Fundamentals: what the KV cache is, and why to treat it as data rather than scratch space
  • Where prefix caching stops, and the open-source stack that goes further: LMCache, CacheGen, CacheBlend
  • The evidence: about 3× faster TTFT at equal quality, and the workloads where reuse pays most
  • Architecture: a layer that drops into vLLM and the stack you already run, managed or self-hosted
  • A 30-second self-check: is KV-cache reuse worth it for you?
Open PDF22 pages · 2.1 MB
Talk

Practical Edge AI: From Runtime Choices to Real-World Product Challenges

Sattaya SingkulTrue Digital Group

Not how to deploy a model once, but how to keep it alive on devices you cannot see. Sattaya Singkul walks the edge-to-cloud spectrum, the fragmented runtime landscape (seven engines, seven hardware envelopes), running an LLM on a phone with LiteRT-LM, speculative decoding, benchmarking on real devices, and partitioning work between edge and cloud with privacy decided by architecture.

In the deck

  • Edge AI in one line, and why "edge" is a spectrum rather than a place
  • The fragmented runtime landscape: seven engines, seven hardware envelopes, no single winner
  • Running an LLM on a phone, honestly: LiteRT-LM and mixed 2/4/8-bit Gemma on device
  • How speculative decoding actually speeds things up
  • Measure, do not guess: benchmark on real devices with AI Edge Portal
  • Edge vs cloud is a spectrum: partition each stage on four axes, privacy by architecture
  • The number you forgot to measure: inference time is not latency, and a demo ends where a product starts
Open PDF20 pages · 2.3 MB
Talk

Local AI for Autonomous Vehicles: From CARLA Decision-Making to Context-Aware and Resilient Urban Mobility

Aueaphum AueawatthanaphisutTensai Thaizen Co., Ltd.

Running the autonomous-driving stack locally: SAE levels and NVIDIA Alpamayo 2 Super, the sensor suite, digital twins, and the open-source CARLA simulator with configurable traffic. Aueaphum Aueawatthanaphisut walks through four decision-making scenarios and the research behind context-aware decisions with LLMs, roadside sensor networks for resilience, and LLM-augmented reinforcement learning for emergency-vehicle prioritisation.

In the deck

  • Autonomous driving today: SAE levels and NVIDIA Alpamayo 2 Super, an open reasoning VLA model for Level 4 driving
  • The sensor suite (camera, radar, LiDAR) and digital twins for automotive systems
  • CARLA: the open-source simulator, its client-server architecture, NPCs and traffic composition
  • Four decision-making scenarios in CARLA: a vehicle running a red light, an obstacle on the road, sudden braking and an emergency lane change, plus pedestrian-crossing prediction
  • Context-aware decision making in autonomous vehicles with social-behaviour modelling and LLMs
  • Resilience: roadside sensor networks for robust perception in challenging environments
  • LLM-augmented reinforcement learning for emergency-vehicle prioritisation under urban traffic stress
Open PDF35 pages · 8.2 MB
Talk

Mamba Inference on MCUs

Friso SmitTU Delft

Can Mamba, the state space model architecture released at the end of 2023, run on a microcontroller with 512 KB of RAM? Friso Smit compares it with transformers, builds a pipeline from PyTorch through ONNX to ESP-DL with Optuna search and int8 quantization, and reports latency and accuracy on an ESP32-S3 for human activity recognition and keyword spotting.

In the deck

  • What Mamba is: from recurrent networks to state space models and selective state space models
  • The target: an ESP32-S3 with a 240 MHz dual core, 512 KB of RAM and 8 MB of flash, about 1000× less than a smartphone
  • Why Mamba: constant memory and linear time, where a transformer’s memory and time grow with sequence length; throughput and scaling comparisons; and hybrid models such as Nemotron 3 and Jamba
  • The research question and challenges: is Mamba suitable for MCUs, how to compare architectures, and training on a PC to run on a chip built for far less than a GPU
  • The toolchain: Optuna hyperparameter search, PyTorch, ONNX and ESP-DL, with bidirectional Mamba, lookup tables and int8 quantization as optimisations
  • Results on human activity recognition and keyword spotting: operator profiling, accuracy against latency, quantization loss and other MCU models
  • Conclusion and future work: Mamba can be very effective on MCUs, but software support and quantization without retraining are both limited; next, better quantization (PTQ or QAT), a dedicated Mamba kernel, and Mamba-2 with comparisons against other models such as transformers
Open PDF34 pages · 6.0 MB

Programme

From the day-of programme. Sessions with slides link to their details above.

  1. 12:30 – 13:00

    Registration

  2. 13:00 – 13:15

    Pie & AI opening

    Andrew Ng

  3. 13:15 – 13:30

    NVIDIA Isaac & Ecosystem Future of the Physical AI

    AGICAFET

  4. 13:30 – 14:00

    Agentic Orchestrate on the Local

    IBM

  5. 14:00 – 14:30

    KV cache Techniques in LMCache

    Tensormesh

  6. 14:40 – 15:00

    Break

  7. 15:00 – 15:30

    Practical Edge AI: From Runtime Choices to Real-World Product Challenges

    True Digital

  8. 15:30 – 16:00

    Local AI for Autonomous Vehicles: From CARLA Decision-Making to Context-Aware and Resilient Urban Mobility

    Aueaphum Aueawatthanaphisut

  9. 16:00 – 16:30

    Mamba Inference on Embedded Systems

    Friso Smit

  10. 16:30 – 17:00

    Let vLLM decide which skill to use from a massive skill collection

    Datakarate

  11. 17:15 – 18:00

    Networking

All times are Bangkok time (ICT, UTC+7).

About the event

An afternoon on AI at the edge, organised by AGICAFET as a DeepLearning.AI Pie & AI meetup at Cleverse Rama 9: an opening from Andrew Ng, then talks on NVIDIA Isaac and physical AI (AGICAFET), agentic orchestration on local hardware (IBM), KV-cache reuse (Tensormesh), practical edge AI (True Digital), local AI for autonomous vehicles (Aueaphum Aueawatthanaphisut), Mamba inference on embedded systems (Friso Smit) and letting vLLM pick the right skill from a large collection (Datakarate), with networking to close.

Featured at this event

  • DeepLearning.AI
  • IBM
  • Tensormesh
  • True Digital
  • Datakarate
  • Cleverse

Stay in the loop

Occasional email about vLLM Bangkok Day and what the Thai vLLM community is building. No spam, unsubscribe any time.