Jul 18, 2026
Rush Flutter to Native Speed — with Rust
Pushing Flutter toward native speed with Rust FFI. A Google I/O Bangkok session on using Rust for the performance-critical paths a Flutter UI cannot reach on its own.
View Slides →Amorn Apichattanakul
On-device AI for mobile, shipped to millions.
I build privacy-first AI that runs directly on device — LiteRT inference, local LLMs, on-device RAG, and biometrics. Right now: face liveness eKYC in a financial super-app serving 4M+ users.
Staff Mobile Engineer at KBTG · Flutter GDE · Deep learning since 2018
Impact & Recognition
On-Device Face Liveness eKYC in Production
4M+ Users on a Financial Super-App
Zero Breaches · Zero Downtime Through Major Platform Shifts
Flutter GDE (Google Developer Expert)
International Tech Speaker (Google I/O & DevFest)
Staff IC · Technical Authority for 60+ Mobile Engineers
SNAPSHOT
At a Glance
Mobile
16 years hands-on IC
Flutter · iOS (Swift) · Android (Kotlin)
Platform Channels · Rust FFI
Staff Mobile Engineer @ KBTG
On-Device & Applied AI
Deep Learning since 2018 · Stanford/Coursera ↗
LiteRT · LiteRT-LM (Gemma) · Gemini Nano
On-Device RAG · Vision + LLM Pipelines
Claude Certified Architect · verify ↗
Banking
Since 2019 at KBTG
Face liveness eKYC · 4M+ users
Security · compliance · zero downtime
Flutter GDE · Dec 2022
FEATURED WORK
Production Impact
On-Device Face Liveness for Regulated eKYC · KBTG
▸↓40% latency
▸↓30% memory
▸Zero network calls during
liveness detection
Regulated eKYC onboarding demands real-time face liveness — confirming a real human, not a photo, video, or mask — running entirely on-device with zero network latency. Once liveness passes, the verified frame is sent for server-side face comparison against the ID photo. Since 2022, I've led the integration and ongoing optimization of LiteRT face liveness: architecting iOS/Android bridges via platform channels, tuning inference across device variance, and coordinating across data science, mobile, and security teams.
Result: production on-device liveness detection running across millions of daily authentications. I published the high-performance Flutter approach behind it (read on Medium) — other banks have since adopted it in their own apps.
Banking-Grade Performance: Biometric Login at 4M-User Scale · KBTG
▸Fastest biometric login in the
industry
▸↓40% launch time
▸↓60% API calls
A financial app serving millions needed enterprise-grade performance under strict transaction security and volatile network conditions across diverse devices. I architected and implemented a high-performance optimistic-loading pipeline — executing flows instantly while resilient fallback callbacks catch transactional anomalies and securely enforce safety policies (logging the user out on critical mismatches). I also optimized the native biometric security loops by eliminating redundant platform-channel context handshakes.
Result: the fastest biometric login speed in the financial industry, with a flawless compliance record and zero security breaches.
Flutter-Native Hybrid Architecture at 60-Engineer Scale · KBTG
▸60+ mobile engineers aligned
▸Zero downtime through major platform
shifts
▸Reference hybrid architecture
Scaling mobile architecture across 60+ engineers within a 3,000-person tech organization while staying hands-on in the codebase. As Staff Mobile Engineer and Head of the Mobile Guild, I lead through code: authoring the reference Flutter-Native Add-to-App bridges, defining Clean Architecture and security standards, and owning the CI/CD and observability pipelines (Firebase, Analytics) that keep a 4M+ user super-app moving fast safely.
Result: unified engineering standards across teams, reduced technical debt, and a "Zero Downtime" rollout strategy that absorbed major platform shifts (e.g., Apple Privacy Manifests) without service interruption.
CURRENT R&D
On the Horizon
What I'm building and benchmarking right now — pushing on-device intelligence further for mobile apps that can't compromise on privacy or latency.
On-Device RAG & Small-Model Fine-Tuning
Building fully local Retrieval-Augmented Generation (RAG) pipelines in Flutter — on-device embeddings, local vector search, and context retrieval for small language models (Gemma / Gemini Nano) with zero cloud round-trips. Next up: fine-tuning lightweight open models for domain-specific mobile tasks.
Predictive × Generative On-Device Pipelines
Wiring fast ~120ms LiteRT detectors (EfficientDet & BlazePose in background isolates) to slow-reasoning local LLMs (Gemma 4 E2B via LiteRT-LM function calling) — turning live camera signals into autonomous offline actions in airplane mode.
Agent-to-UI (A2UI) with GenUI & Rust FFI
Streaming adaptive widgets dynamically from AI agents in Flutter instead of static screens, while offloading compute-heavy image and data bottlenecks to native speed via Rust FFI.
ABOUT AMORN
Three Long Tracks.
Three Long Tracks.
One Intersection.
I've been a hands-on mobile engineer for 16 years, since the iPhone 3GS. I began studying deep learning in 2018 (Deep Learning Specialization, Stanford/Coursera) — well before the on-device AI wave. And since 2019 I've built inside the hardest proving ground I could find: financial banking at KBTG, where I've deliberately stayed on the Staff individual contributor (IC) track — writing production code every day while guiding technical architecture for 60+ mobile engineers in a 3,000-person tech organization. Those three tracks now converge in production: LiteRT face liveness eKYC serving 4M+ users.
My work sits where mobile, on-device AI, and banking meet — not as a job pivot, but as a deliberate, multi-year craft. As a Staff IC, I lead by building the hardest pieces first — native bridges, edge inference pipelines, and performance benchmarks — and turning what works into the architectural standard. Three principles guide that work:
- Mobile depth where it counts. Deep expertise in Flutter and iOS (Swift), working Android (Kotlin) experience, and Rust FFI when Dart hits a native ceiling — picking the right tool for the constraint, not by preference.
- Applied AI in production, not in demos. On-device LiteRT detectors, local LLMs (LiteRT-LM / Gemma, Gemini Nano), and on-device RAG — shipped, profiled, and tuned for real device variance while keeping privacy on the device.
- Banking-grade engineering. Security, compliance (Privacy Manifests), and zero-downtime under regulatory scrutiny — proving architectural standards in production code for systems regulators audit and 4M people rely on every day.
This is the work I've been quietly preparing for. If it intersects with what you're building, I'm always up for a conversation.
THE NICHE
Why Banking Is the Proving Ground
On-device AI is easy to demo and hard to ship. Banking is where the constraints are real.
Spoofing must be caught before any network call
On-device LiteRT face liveness confirms a real human — not a photo, video, or mask — entirely on-device, with zero network round-trips during detection. Only after liveness passes does the verified frame go to the server for face comparison against the ID photo.
Onboarding is regulated (eKYC)
I designed a high-performance Flutter eKYC approach and published it — other banks have since adopted it in their own apps.
How I built it — Medium ↗The environment is adversarial
Anomaly-triggered enforcement policies, zero security breaches, and a flawless compliance record at 4M+ user scale.
Downtime is not an option
A zero-downtime strategy that absorbed major platform shifts (Privacy Manifests, OS updates) — governed for 60+ engineers.
SKILLS
Core Competencies
On-Device & Applied AI
Mobile Systems & Native Performance
Identity, Biometrics & Banking Security
COMMUNITY & EVANGELISM
Sharing Knowledge at Scale
Disseminating production-scale mobile and AI engineering insights to the global developer community.
3.5+ yrs
as Flutter GDE
24+
talks delivered
42+
articles on Medium
160K+
reads on Medium
Google I/O Connect
GDG DevFest
Build with AI
Tech Campus
Flutter Mekong
Featured Talk · Jul 19, 2026 · Google Cloud Next Extended
Predictive × Generative: From Signal to Meaning, On-Device
Fast predictive vision models (~120ms/frame EfficientDet in an isolate) and local LLMs (Gemma 4 E2B via LiteRT-LM function calling) working together on-device — turning live camera signals into meaning and action in airplane mode, with zero cloud round-trips.
More Talks
Nov 1, 2025
On-Device AI with LiteRT
Exploring on-device AI as a complement to cloud AI — practical mobile implementation using LiteRT (TensorFlow Lite) for low-latency, privacy-preserving features.
View Slides → Watch on YouTube →
Feb 18, 2026
A2UI: GenUI with Flutter
Exploring Agent-to-UI (A2UI) architectures using dynamic Generative UI (GenUI) in Flutter. AI agents that stream, construct, and render customized UI widgets in real-time.
View Slides → Watch on YouTube →
Nov 23, 2025
Flutter & Flash: AI Mascot on the Edge
On-device AI using Gemini Nano to create an AI-powered mascot on edge devices. Flutter integration with on-device AI for personalized experiences without server dependencies.
View Slides →
Sep 16, 2026
Smarter with AI Agents
A hands-on session at Techtopia on working with AI agents — setting them up and putting them to work on real developer tasks.
Jul 26, 2025
The Power Couple: Flutter & Native
Why choose between native and Flutter when you can use both? Flutter's Add-to-App feature for seamlessly integrating Flutter modules into existing iOS and Android projects.
View Slides →Selected Articles from Medium
Aug 4, 2026
On-Device AI Needs Both: the AI That Sees and the AI That Thinks
A fast detector watches every frame. A local LLM reads what changed and decides what it means. Together they catch a phone thief on-device, in airplane mode.
Read on Medium →Jul 20, 2026
Flutter Too Slow for Images? Rush It with Rust
Flutter developers often blame images for a slow app. A real-device benchmark shows the gap is the codec, not Dart — and a small Rust crate closes it.
Read on Medium →Mar 24, 2026
One Size Fits None: Adaptive, Context-Aware UI in Flutter with GenUI and A2UI
In 2023, when Gemini was first released, we were amazed AI could answer questions we couldn't even find on Google. By 2026, we've grown tired of just reading.
Read on Medium →
CONTACT
Let's Connect
Open to conversations about mobile, on-device AI, and banking-grade engineering.