Sasha Denisov
EPAM, Chief Software Engineer, AI, Flutter, Dart and Firebase GDE
Berlin, Germany
Actions
Sasha is a software developer with over 20 years in multiple technologies and domains. A tech lead, architect, and mobile solutions expert, he specializes in AI integrations for mobile apps, focusing on on-device AI. He excels in Flutter, Firebase, and generative AI for robust, scalable apps. Sasha is Chief Software Engineer and Head of Flutter Discipline at EPAM, GDE for AI, Firebase, Flutter, Dart, and co-organizes Flutter Berlin Community
Badges
Area of Expertise
Topics
Smart Routing, Simple Flows: An Alternative to Complex Agent
When we started building our multi-brand AI search assistant, the industry buzz pointed us toward autonomous agents. We evaluated the popular frameworks — LangGraph, CrewAI, AutoGen, and others — each promising flexible, self-directing AI systems. But as we dug deeper, concerns emerged: unpredictable latency as agents decided their own paths, token costs multiplying with each reasoning loop, and debugging sessions that felt like archaeology.
We stepped back and asked: what do we actually need? Our assistant handles product search, recommendations, and customer questions across multiple retail brands. The tasks are complex but well-defined. Users expect fast, consistent responses. We need to test, monitor, and explain system behavior.
The answer wasn't agents — it was smart routing with simple flows. We built our system on Genkit with a single architectural principle: classify intent once, then execute deterministic flows. One LLM call decides the path (product search, clarification, comparison, off-topic). From there, specialized flows handle RAG retrieval, filter extraction, and streaming response generation — all predictable, all testable.
The results speak for themselves: sub-second routing decisions, consistent response times, straightforward debugging, and a codebase that new team members can understand in days, not weeks. We now serve multiple enterprise clients across retail and cosmetics, with the same architecture handling different brands, catalogs, and deployment targets.
This talk shares our evaluation journey, the "smart routing, simple flows" architecture we landed on, concrete implementation patterns with Genkit, and honest lessons about when this approach works — and when you might actually need agents.
Mobile and Browser AI Agents Without the Backend
Building AI agents no longer means the cloud is mandatory. AI agents can now run directly in the browser and on mobile devices. This talk explores "on-device" processing (Edge AI), which allows you to build intelligent agents that operate entirely offline. You no longer need a constant internet connection, and you stop paying for every API call. This unlocks new functionalities, ensures data privacy, and provides zero latency.
We'll focus on lightweight (meaning compact enough to run on-device) open models like Gemma, Llama, Deepseek, and others. They are designed to run efficiently on client devices and serve as the "brain" for our offline agents.
This talk is a practical guide to building these systems. We will show how to implement full-fledged agent capabilities, not just simple inference:
On-device Function Calls: How to teach the model to interact with device APIs (e.g., reading contacts or calendar events) without an internet connection.
On-device RAG: How the agent can query local data (documents, notes, emails) to provide highly relevant, context-aware answers.
We will cover the architecture and practical steps for integrating these models in both a browser and a mobile application, and frankly discuss the trade-offs: weighing the benefits (privacy, offline) against the challenges (model size, performance).
Hosting an AI Debate Club
Welcome to the AI Debate Club! In this workshop, you'll build a Flutter app that hosts lively, autonomous debates between different AI models.
First, we'll construct the debate arena. You’ll learn how to bring AI debaters into the Flutter ring using both cloud and on-device approaches. We'll cover powerful cloud contenders via the Google SDK, Firebase Vertex AI, and REST APIs, as well as on-device champions like Gemma and Mistral Small that enable lightning-fast, offline debates using Edge AI.
Next, we’ll scout the talent. We’ll analyze top AI models—like Gemini, ChatGPT, and Claude—comparing their unique debating styles (capabilities) and membership fees (pricing) to determine who argues best for your app's needs.
For the main event, we'll build the app live. You will witness multiple AIs autonomously interact, respond, and "argue" their points, all orchestrated within Flutter with no human intervention required once the debate begins.
You'll leave fully equipped to host your own AI debates, ready to deliver dynamic, cutting-edge, and maybe even slightly chaotic AI experiences to your users.
Future of Mobile AI. What On-Device Intelligence Means for App Developers
Two years ago, adding AI to your app meant only cloud APIs — sending data to servers, paying per request, hoping for good internet. That world is ending.
Today, you can run SLM directly on a phone. No internet, no per-request costs, data never leaves the device. This is production-ready technology.
I built flutter_gemma, an open-source plugin for running AI models locally on iOS, Android, and Web. I'll share what I've learned—not marketing, but real trade-offs and opportunities.
What's possible now — Running Gemma 3 on smartphones, the hardware that enables it, formats that matter (.task, .litertlm). What changes for developers — Offline-first AI, hybrid cloud/edge patterns, model size decisions, fine-tuning and optimization skills.
Where we're heading — Multimodal on-device models, function calling, on-device fine-tuning, edge-specific models like Gemma 3n.
The future isn't replacing cloud — it's a new option: private, fast, and works anywhere.
Building Voice AI Agent That Listens, Understands, and (Most Importantly) Sells
Voice commerce is evolving beyond simple chatbots. Google's Multimodal Live API brings 600ms-latency real-time conversations with native audio processing, multimodal understanding, and natural interruption support through bidirectional WebSocket streaming.
We built an AI shopping assistant that doesn't just talk—it understands products, answers technical questions, and guides customers through purchases using voice. By combining Live API with RAG, our assistant accesses product catalogs and knowledge bases in real-time through function calling, creating seamless consultation experiences.
This talk walks through our development journey from prototype to production. You'll see how we integrated Live API's multimodal capabilities with RAG infrastructure, designed function calling patterns for dynamic product queries, managed conversation context across sessions, and optimized for real-time performance.
We'll share practical lessons: WebSocket architecture decisions, RAG pipeline design for product knowledge, handling edge cases in voice interactions, and scaling considerations. You'll leave with actionable insights for building voice-enabled applications understanding when RAG enhances conversational AI, and avoiding common pitfalls we encountered.
Perfect for developers exploring voice commerce, conversational AI, or RAG implementations in production systems.
AI Agent Specialization. RAG vs Fine-tuning
Many AI agents can be built successfully using just a base model and good prompting. But what happens when that's not enough? How do you make an agent understand your internal context, your proprietary data, or speak in your unique style?
This is where specialization comes in. This talk is a deep dive into the two fundamental strategies for it: RAG (Retrieval-Augmented Generation) and Fine-tuning.
RAG (Retrieval-Augmented Generation): We will explore how to implement a RAG pipeline. This means setting up a vector database, search, and context retrieval. This is the best method for giving an agent access to dynamic, easily updated external knowledge (like company docs or user data) without modifying the model itself.
Fine-tuning: Next, we will analyze Fine-tuning. This method "bakes" knowledge into the model, making it static. It is also the primary way to change the model's fundamental behavior and style (e.g., teaching it your tone of voice or a specific response format).
The central question of the talk is: RAG or Fine-tuning? We won't just compare them; we'll provide a clear decision framework:
When to use RAG for dynamic, easily updatable knowledge?
When to use Fine-tuning for static facts or core behavioral skills?
And how (and why) to combine both approaches to create the ultimate specialized AI agent.
Attendees will leave with a practical guide to choosing the right specialization strategy for their AI application.
Make Your AI Agent's Path to Production Shorter
For developers, architects, and ML specialists moving AI concepts into production, building a complete, agent-based product is often confusing and complex. The landscape is saturated with new AI frameworks, many of which reinvent their own solution architectures. This creates confusion about the "right way" to build, making even seemingly simple applications look overly engineered. This talk charts a pragmatic path through this complexity using solutions from Firebase.
We will tackle this challenge from two angles using Firebase AI Logic and Firebase Genkit. AI Logic enables rapid AI feature development via secure, client-side SDKs with no backend required. Genkit, in contrast, is an open-source, server-side framework for building complex, observable, and reliable AI flows.
In this session, we'll break down how to:
Build a Flexible Architecture: Show how AI Logic solves the "quick demo" problem and how Genkit solves the "reliable product" problem (RAG, agents).
Ensure Reliability and Security: Discuss the built-in approaches to security (key and API protection), testing, and monitoring that simplify maintenance.
Enable Simple Evolution: Explain when to use the client-side approach, when to use the server-side one, and how they can coexist, allowing an application to scale without technical debt.
This talk is for those who are tired of the hype and are looking for the most direct path to learning how to build production-ready AI applications, AI agents, and the right architecture for real products that just work.
Hosting an AI Debate Club
Welcome to the most exclusive club in tech: The AI Debate Club, hosted right inside your Flutter app! This workshop is your all-access pass to setting up lively (and autonomous) debates between different AI models. Forget boring integrations; we're building an arena for intelligent discourse, moderated only by code!
First, we'll construct the debate stage. Discover all the ways to bring your AI debaters into the Flutter ring: leveraging the Google SDK, Firebase Vertex AI, Firebase Genkit, and versatile REST APIs for powerful cloud-based contenders. We'll also explore setting up ringside seats using Edge AI with on-device models like Gemma, Deepseek, and Mistral Small, allowing debates to happen directly within the app, lightning-fast and even offline!
No debate club is complete without its opinionated members! We’ll invite today's top AI debaters like Gemini, ChatGPT, Gemma, and Claude to join. We’ll analyze their unique debating styles (capabilities) and potential membership fees (pricing) to figure out who argues best for your specific app's needs and topics.
For the grand finale – the main event! Together, we'll build and unleash the ultimate AI Debate Club application. Witness multiple AI models autonomously interact, respond to each other, and perhaps even "argue" their points – all orchestrated within Flutter, with no human intervention required once the debate begins!
Expect plenty of fun and 'aha!' moments as we push the boundaries of AI collaboration in Flutter. You'll leave not just inspired, but fully equipped to host your own AI debates and bring truly dynamic, cutting-edge, and maybe even slightly chaotic AI experiences to your users.
Building Offline AI Agent in your Flutter app
Discover how to build the next generation of mobile apps with Offline AI Agents. This talk explores leveraging Edge AI and on-device processing to create intelligent agents that operate entirely without an internet connection, enhancing privacy, reducing latency, and enabling new functionalities independent of the cloud.
We'll focus on using lightweight, open models like Google's Gemma family, designed to run efficiently on mobile devices. These models form the "brain" of our offline agents. A key part of this session introduces a custom Flutter plugin that seamlessly integrates these models, making it significantly easier to build offline AI capabilities into your Flutter projects.
Learn the practical steps for implementation, from selecting on-device models (e.g., Gemma, Deepseek, Llama) to integrating them with our Flutter plugin. We will discuss the trade-offs of Edge AI—weighing benefits like offline access and privacy against challenges like model size and device performance compared to cloud AI. We will also explore techniques for structuring agent tasks, fine-tuning models, and delving into on-device Retrieval-Augmented Generation (RAG). This advanced RAG capability enables agents to query local data sources for highly relevant, context-aware interactions, entirely offline.
Serverless backend with Firebase Cloud Functions and beyond
In this talk, I will describe how to build an event-driven serverless backend using Firebase services and Cloud Functions and show examples, of course. Did you know, for example, that last year a new version of Cloud Functions, v2.0, was released, which now operates fundamentally differently, opening up many new opportunities? Or that a cloud function can be written not only in JS and TS but also in Python or Java, and that you can create your own event that will trigger the function? Or that you can connect any Firebase service, or any Cloud API, for example, Genkti, or an external API, such as ChatGPT? That eventually, you can create your own function in a language not natively supported out of the box, like Dart or Rust. Let's figure it out together, how it works!
Flutter and AppleTV. Story of one challenge
Flutter officially supports six platforms, but, unfortunately, TV platforms are not yet on the list. While Flutter applications can indeed be developed for TV platforms, and launching your app on Android TV can be relatively easy, the situation with Apple TV is far more complex. In this talk, I will describe step-by-step how we solved this problem and finally launched a Flutter application on AppleTV
Let's shake the Flutter tree
The more platforms a Flutter application supports, the more code a project can have that is needed only for one platform and not needed for others. How to make sure that unnecessary code does not get into the final platform-specific build? And how to check that it really did not get?
In Flutter, there is a Tree-shaking mechanism, which is designed exactly for these purposes. Many Flutter developers have heard about it, but not everyone really understands what it is and how it actually works. Let's try to figure it out.
Workshop: Flutter app Journey from Mobile to Web and Desktop
Flutter is a great multi-platform framework, that provides an opportunity to create not only mobile but also web and desktop apps, but web and desktop applications have different UI/UX in comparison with Mobile, at least because they need keyboard and mouse control. It is completely different than working with a touchscreen.
In this workshop, I'm going to show, how to remake a mobile app into a web/desktop one and figure out how to properly manage focus and what a hover effect is.
Workshop: Flutter for Android TV and Apple TV
Get ready to explore TV app development using Flutter. While Flutter officially supports various platforms, TV platforms like Android TV and Apple TV have been overlooked. But fear not! In this workshop, we'll show you how to bring your Flutter apps to the big screen.
In this workshop, we will go through all the steps of porting your mobile app to the big screen, discuss all the nuances of TV apps, such as controlling the app with a remote control, and finally launch the app on a TV platform, not only on Android TV, but also on Apple TV
Optimization of Mobile Development Strategy for Maximum Business Impact"
In this presentation, we'll explore practical strategies for optimizing mobile development to achieve exceptional results. We'll delve into the nuances of technology selection, comparing options like native iOS-Android, hybrid frameworks, Flutter, React Native, Kotlin Multiplatform and so on. Additionally, we'll discuss structuring mobile delivery processes, effective team organization, role assignments, process refinement, and leveraging appropriate tools. Furthermore, we'll examine the potential expansion of mobile solutions to web, TV, and desktop platforms, along with insights on smooth platform migration for enhanced optimization. Join us as we uncover actionable insights to elevate your mobile development endeavors to new heights.
Workshop: Bring Your Fine-Tuned AI Model On-Device with Flutter
In this hands-on workshop, we’ll explore the groundbreaking family of open AI models by Google: Gemma. You’ll learn how to leverage the power of Gemma to build innovative Mobile and Web projects that run directly on-device or in the browser, without requiring an internet connection.
Together we are fine-tune the Gemma model with your data and integrate it into your Flutter app for running locally without internet connection
What you'll gain from this workshop:
- A deep understanding of Gemma's core functionalities and lightweight, state-of-the-art architectures.
- Insights into the benefits of using Gemma, including its open-source nature, versatility across tasks, and ability to run on-device or in the browser.
- Practical experience with fine-tuning Gemma models to achieve better results across a wide variety of tasks. Once fine-tuned, models require fewer examples in the prompt to perform well.
- Step-by-step guidance on how to integrate Gemma into your Mobile and Web applications, unlocking new possibilities for user interaction and app functionality.
-An overview of different Gemma models (CodeGemma, PaliGemma, etc.), their use cases, differences from Gemini, and tips on selecting the right model for your specific project needs.
AI in Your App with Firebase: AI Logic vs. Genkit
Firebase provides developers with a powerful suite of tools for integrating generative AI, but choosing the right architecture is critical. This talk presents a direct comparison of two primary approaches: using the Firebase AI Logic client SDKs versus implementing the Genkit backend framework.
We will begin with Firebase AI Logic, which provides secure SDKs for calling Gemini models directly from the client device (web, iOS, or Android). We'll explore its most advanced features, including the Gemini Live API for real-time, bidirectional voice and text streaming, enabling truly natural, conversational experiences. Critically, we will also cover its "hybrid AI" strategy, which intelligently blends fast on-device models (for privacy, offline access, and speed) with more powerful cloud-hosted models, ensuring your app's AI features are both responsive and reliable.
Next, we will dive deep into Genkit, an open-source framework for building, deploying, and monitoring robust AI flows on the backend. Genkit is designed for complex, server-side tasks:
Retrieval-Augmented Generation (RAG): Connecting AI to your own data in Firestore or other sources.
Multi-step "Flows": Orchestrating multiple model calls, APIs, or database lookups.
Tooling: A local UI for testing, tracing, and observability.
Deployment: Seamless integration with Firebase Functions and App Hosting.
This session will feature practical code examples for both methods and will clearly define the decision point: when is the client-side power of AI Logic sufficient, and at what point should you adopt Genkit to build scalable, testable, and production-ready AI backends?
Social Developers Conference 2024 Sessionize Event
Open Source Day 2024 Sessionize Event
DevFest Venezia 2023 Sessionize Event
Devfest Pescara 2023 Sessionize Event
droidcon Uganda 2023 Sessionize Event
droidcon Kenya 2023 Sessionize Event
DevFest Hamburg 2023 Sessionize Event
DevFest Armenia 2023 - 10th Anniversary Sessionize Event
Devfest Georgia 2023 Sessionize Event
droidcon Lisbon 2023 Sessionize Event
Flutter Firebase Festival Sessionize Event
Flutter Wrocław Meetup Sessionize Event
AndroidMakers by droidcon 2023 Sessionize Event
DevFest Pisa 2023 Sessionize Event
Flutter Heroes 2023 Sessionize Event
DevFest 2022 | Gothenburg, Sweden Sessionize Event
droidcon Egypt 2022 Sessionize Event
droidcon Berlin 2021 Sessionize Event
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top