Ashish Mehrotra

Exploring on-device LLMs, Edge AI & agentic systems for mobile.
Building, learning and sharing along the way.

[Ashish Mehrotra @edgewithashish]

AI has been writing one word at a time. Google just changed that.

DiffusionGemma refines an entire 256-token block together instead of generating left to right. 1,000+ tokens/sec on an NVIDIA H100.

I built a live demo to see how it compares with regular Gemma 4.#DiffusionGemma

[Ashish Mehrotra @edgewithashish]

@maxzrco Let’s connect

[Ashish Mehrotra @edgewithashish]

HTTP standardized how applications communicate.

A2A protocol is trying to do the same for AI agents.

Built a Codebase Explainer with ADK, LangGraph, and CrewAI to learn how it works in practice.

Video, code, and blog - https://t.co/njLO4MzdJ3

#A2A #AIAgents #AIEngineering

[Ashish Mehrotra @edgewithashish]

Full blog: https://t.co/X3VDIpKdEi

[Ashish Mehrotra @edgewithashish]

"Skip the infrastructure. Ship the agent."

One of the Google I/O 2026 announcements that immediately caught my attention was Managed Agents.

Define the agent. Google handles the runtime environment around it.

Wrote a deep dive 👇

#GoogleIO2026 #AIAgents #ManagedAgent

[Ashish Mehrotra @edgewithashish]

@GoogleAIStudio Built “Reading DNA” - A Telegram AI agent that learns your reading taste through conversation and recommends books using ADK 2.0+ Gemini 2.5 Flash.
#GoogleADK #AIAgents

[Ashish Mehrotra @edgewithashish]

Project + Architecture : Reading DNA is a Telegram-based AI agent built with Google ADK 2.1, FastAPI, Gemini 2.5 Flash, Vertex AI, and Cloud Run.
It analyzes reading taste through multi-turn conversations and recommends books with contextual reasoning.
https://t.co/3YVUAAo0s3

[Ashish Mehrotra @edgewithashish]

Google ADK 2.0’s shift toward graph-based workflows caught my attention at #GoogleIO2026.
Built “Reading DNA” while exploring it- A Telegram AI agent that learns your reading taste through conversation and recommends books using ADK + Gemini 2.5 Flash.Demo 👇
#GoogleADK #AIAgents

[Ashish Mehrotra @edgewithashish]

@googleaidevs Built an offline multimodal recipe analyzer with Gemma 4.Scan recipes in any language the app translates them into English and converts them into structured recipes fully on-device. No internet connection needed.

[Ashish Mehrotra @edgewithashish]

Built an offline multimodal recipe analyzer with Gemma 4. Scan recipes in any language the app translates them into English and converts them into structured recipes fully on-device. No internet connection needed.
@GoogleAI @AndroidDev

#Gemma4 #OnDeviceAI #AI #GoogleIO