DevFest Bucharest

‍

‍

‍

-

Firebase is a staple of Android development but do you know everything about Firebase and the libraries that belongs to the ecosystem?

This session will take you beyond the basics and show you the hidden gems of the Firebase ecosystem. We'll explore how to use Firebase Performance Monitoring to identify and resolve performance issues, implement robust security measures with App Check, and dynamically tailor your app's behavior with Remote Config and much more!

Learn how these powerful features can help you build faster, more secure, and more engaging Android experiences.

‍

The Vibe Effect talk introduces a modern workflow that uses AI prompting to draft video effects in seconds. You will learn how to bridge the gap between creative vision and technical execution by using the Gemini SDK to generate shader code, then applying it to video files using Jetpack Media3 editing APIs.

In 2026, "it has AI" is the new standard. From Gemini Nano to custom LiteRT models, we are pushing incredible intelligence to the edge. However, this power comes with a hidden cost: the "AI tax" that we all experienced. The phone heats up, the frame rate throttles, and the battery percentage freefalls. As developers, we focus out attention on inference latency and model accuracy, but we often ignore the metric that actually drives uninstalls: energy efficiency.

This session treats energy as a key point of the modern Android AI stack. We move beyond basic optimizations and dive deep into the 2026 updates to the Android Studio Power Profiler, demonstrating precisely how to isolate and visualize the power consumption of NPU inference versus CPU fallback. We will learn how to spot "energy leaks" where models keep the hardware awake longer than necessary or fail to delegate properly to the NPU.

Beyond profiling, we will look at how we can implement a "thermal-aware" architecture. Using Android's Thermal API and the Android Dynamic Performance Framework, we will learn how to build apps that dynamically regulate their own "AI metabolism". We will look at how we can detect when a device is under thermal stress and make it automatically downgrade to a lighter quantized model or defer to background inference tasks.

Whether you are building generative agents or real-time computer vision, you will leave this talk with the tools to audit your app's energy footprint and a concrete strategy to deliver powerful AI features that respect your user's battery life.

-

-

-

-

‍

‍

‍

-

‍

Moving beyond theoretical discussions, this talk focuses on the practical steps and frameworks necessary to build responsible AI. We will explore actionable strategies for addressing bias, promoting fairness, ensuring transparency, and ultimately prioritizing people and ethical considerations in the creation and deployment of data-driven technologies.

In this presentation I will show how I evolved a simple medical chatbot application into an agentic healthcare assistant, combining multi-modal SLM MedGemma with Gemini to process and interpret not only medical images but medical records and reports in various formats, transform medical transcripts in SOAP format, while leveraging ADK and VertexAI, implementing guardrails, orchestration and validations.
The development process is accelerated using Antigravity.

‍

-

-

-

-

‍

‍

‍

AI agents are increasingly being introduced into engineering and security workflows — from code writing, review to analysis and vulnerability triage to infrastructure changes and operational support. As their autonomy grows, so does a practical question many teams are already facing: how much trust should we place in these systems as collaborators?

This talk looks at AI agents through a security and engineering lens, focusing on trust as something that must be designed and verified, not assumed. Based on real-world experience building and operating security programs at scale, I’ll examine where agents are genuinely useful today, where they fall short, and the kinds of risks they introduce when they interact with production systems, sensitive data, or decision-making processes.

We’ll walk through some concrete examples and explain how these translate into operational and security concerns.

The session will conclude with a practical set of principles for integrating agents responsible.

With a strong background in offensive security and vulnerability research, Cosmin focuses on embedding security directly into CI/CD pipelines, cloud environments, and development workflows, transforming security from a gate into an engineering capability.

He is passionate about measurable outcomes, risk-driven vulnerability management, and AI-assisted security workflows. A frequent conference speaker and OSCP-certified professional, Cosmin is equally committed to growing high-performing security teams, fostering a strong research culture, and strengthening customer trust at scale.

‍

In today's rapidly evolving AI landscape, ensuring the security and integrity of your AI systems is no longer optional – it's a critical imperative. But how do you move beyond theoretical discussions and implement robust, continuous security for your AI deployments?

Whether you're an AI developer, a security architect, a DevOps engineer, or simply passionate about building secure AI, this session will introduce you to he SAIF (Security AI Framework) and features a live exploration of the SAIF Map, including the new Version 2.0 for Secure Agents, helping you navigate the complex threat landscape of autonomous systems.

We will then pivot to defense, demonstrating Model Armor in action within a sample application to show how you can enhance safety without sacrificing performance. Finally, we will review the actual deliverables of an AI Security Posture Review, demystifying the audit process.

‍

Why this matters: LLMs and AI agents are everywhere—but so are new attack surfaces. OWASP’s Top 10 for LLM Applications is the emerging standard for securing these systems. This talk translates those risks into actionable defenses, backed by real enterprise experience from Adobe’s Security AI Guild.

What you’ll learn: OWASP Top 10 for AI in practice | Prompt Injection & Indirect Injection | Sensitive Information Disclosure | Model Theft & Data Poisoning | Insecure Plugin/Tool Integration |Excessive Agency & Overprivileged Actions Enterprise guardrails: patterns from the Security AI Guild—retrieval hygiene, policy-driven tool use, safety routers, and multi-layer isolation. Threat modeling for agents: capability scoping, least privilege, and auditability.

Demo Title: From Exploit to Resilience: Guild Patterns that Mitigate OWASP Top 10 for AI

Takeaways: SherIOCk, an open source tool OWASP Top 10 cheat sheet for AI developers A printable checklist and reference threat models

Who should attend: Developers, security engineers, and tech leads building LLM features, agents, or RAG systems.

Commercial spam filters are often “black boxes”. Can a developer build a better solution with zero infrastructure?

In this intermediate-level session, I present a custom, event-driven phishing detection framework built entirely on the Google Cloud Platform. We will move beyond high-level concepts and dissect a production-ready architecture where Cloud Run Functions act as the nervous system, Pub/Sub manages the asynchronous data flow, and the Natural Language API provides the intelligence to analyse semantic intent.

This talk is backed by real-world data. I will share the results of this experiment, comparing the “easiness” of the serverless approach against the detection rates achieved. You will see exactly how to chain these services to create a scalable, private security bot that costs pennies to run but delivers impressive results.

AI agents are no longer passive tools. They reason, act, call APIs, and operate with increasing autonomy, often with far more trust than they deserve. Yet most agent implementations today are built without clear security rules, threat models, or enforcement boundaries.

Without revealing a checklist, this talk exposes why many AI agents would fail a basic security review, and why traditional application security assumptions break down when decision-making is delegated to a model. The focus is not on hypothetical dangers, but on structural weaknesses that repeatedly appear in production systems.

This session is for builders, security engineers, and decision-makers who want to move beyond “it works” and start asking the more important question: what happens when it doesn’t?

Adversaries Don't Operate in Silos - Neither Should Your Detection Strategy

Modern cloud savvy adversaries move fast - often achieving their objectives in less than an hour. In this session, we'll dissect real-world attack campaigns from sophisticated nation-state actors to eCrime groups to understand hoe they operate on cloud infrastructure to achieve their objectives ranging from data exfiltration and cryptomining to distructive actions taken at scale.

What makes these adversaries particularly dangerous is their understanding of the complete cloud ecosystem. Some begin their attacks through traditional entry points - compromised credentials, vulnerable public facing applications, or misconfigurations- while others exploit the software supply chain, leveraging poisoned dependencies or compromised container images. Regardless of the entry point, successful adversaries share a common trait: they move fluidly across multiple layers of the compute stack, from hypervisors to traditional systems, containers or orchestrators.

Through detailed kill chain analysis we wil examine the average execution time at each attack stage, demonstrating a critical challenge: when security teams view these attacks through isolated lenses - monitoring container separateky from VMs, or runtime separately from build time - they only see a fragment of the malicious activity. Adversaries thrive in these gaps, outpacing fragmented defenses at every turn.

Correlation at detection time is the equalizer! By weaving together signals from multiple telemetry sources security teams can reconstruct the attacks as they unfold. This holistic view collapses Mean Time To Respond (MTTR), enabling teams to intercept threats mid attack rather than discovering breaches after the damage is done.

Attendees will learn how modern adversaries weoponize cloud complexity, why detection silos guarantee failure, and how to build security strategies that operate at adversary speed.

We've spent years building CI/CD pipelines, automating infrastructure with Terraform, and pushing for everything-as-code. Most teams have solved the deployment problem. But scaling operations across multiple teams, services, and environments? That's a different challenge.
Platform engineering addresses this by turning operational knowledge into self-service tooling, golden paths, internal developer portals, and standardized workflows that reduce the back-and-forth between dev and ops.
AI is now becoming part of this stack. Not as a magic fix, but as another tool in the toolbox. Smarter alerting that cuts through noise. Automated root cause analysis during incidents. Infrastructure recommendations based on actual usage patterns.
In this session, I'll break down how DevOps practices are evolving from pipeline-centric workflows to platform-driven approaches, where AI integration makes sense today and where it's still maturing, and lessons learned from real-world implementation in cloud environments.
If you've ever spent a weekend debugging an incident that better observability could have caught, or dealt with the overhead of onboarding teams onto inconsistent infrastructure - this talk is for you.

‍