GDG Lisbon Meetup - AI on Android: From Prototype to Production
Details
Everyone is demoing AI on mobile, but getting LLMs into production, whether at scale in fintech or entirely offline on-device comes with challenges the official docs rarely mention.
Join us on August 27th at Critical Software for an evening focused on practical engineering, real benchmarks, architectural tradeoffs, and networking with fellow developers.
## 🗓️ Agenda
- 18:30 – 19:00 | Welcome & Networking
- 19:00 – 19:45 | Talk 1: Bringing LLMs into Production Android Apps (Andrii Veremiienko) + Q&A
- 19:45 – 19:55 | Short Break
- 19:55 – 20:35 | Talk 2: Constraints of On-Device AI (Dmytro Samoilov) + Q&A
- 20:35 – 21:00 | Pizzas, beer and Networking!
## 🎙️ Talks & Speakers
### 1. Bringing LLMs into Production Android Apps: What Actually Ships
Speaker: Andrii Veremiienko
Everyone's demoing AI on Android, but getting an LLM into a real, shipping app is a different story. Andrii will walk through the practical side—on-device vs. cloud tradeoffs, how to architect AI features so they stay fast, private, and maintainable, and where things tend to break in production.
Expect concrete examples from building AI features inside a large fintech app and from shipping his own AI-powered Android apps solo, with takeaways you can use the next day.
> About Andrii:
> Android engineer at Revolut, working on UI and architecture for the Credit Cards vertical and integrating AI features as the product scales internationally. He writes for ProAndroidDev and ITnext, mentors at Women in Big Data and Women Coding Community, and builds his own AI-powered Android apps on the side.
### 2. Constraints of On-Device AI — What I Hit Shipping a 100% Local Note-Taking App
Speaker: Dmytro Samoilov
A couple of weeks ago, Dmytro shipped Offhand—an Android app that records voice memos and turns them into structured notes. Whisper does the speech, Gemma does the structure, and nothing leaves the phone. No cloud, no API key, no subscription.
Getting there meant hitting every constraint the docs don't mention: GPU quirks, thermal throttling that makes iteration five 47% slower than iteration one, a 70-second freeze just to load a model on Pixel 8, and finding a 375 MB model that beat one 7× its size.
Dmytro will share real performance metrics measured across three devices (a flagship, a Tensor Pixel, and a 2020 mid-ranger) and explain why the fight is worth it: on-device inference costs $0 per request, completely changing what you can afford to ship.
## 📍 Location & Logistics
- Venue: Critical Software Offices
- Map Link: Google Maps Location
- Date: August 27
- Time: 18:30 – 21:00
> ⚠️ Note: Spaces are limited. Please keep your RSVP updated so we can accommodate everyone on the waitlist.