Ivan Nardini
@ivnardini.bsky.social
DevRel Engineer @googlecloud in 🌉
In love with minds, souls, Rock 🤟🏻 and Photography 📷
My religion is LARMATS.
In love with minds, souls, Rock 🤟🏻 and Photography 📷
My religion is LARMATS.
🚀 Gemini Tuning updates: Custom metrics and Preference data prep
Many asked, "How do I verify Gemini SFT works?" and "How do I prepare DPO data?"
With Vertex AI ENG team we published tutorials on preparing data and using custom metrics for evaluating tuning with Gemini 2.5 models.
Code in 🧵
Many asked, "How do I verify Gemini SFT works?" and "How do I prepare DPO data?"
With Vertex AI ENG team we published tutorials on preparing data and using custom metrics for evaluating tuning with Gemini 2.5 models.
Code in 🧵
December 4, 2025 at 3:30 PM
🚀 Gemini Tuning updates: Custom metrics and Preference data prep
Many asked, "How do I verify Gemini SFT works?" and "How do I prepare DPO data?"
With Vertex AI ENG team we published tutorials on preparing data and using custom metrics for evaluating tuning with Gemini 2.5 models.
Code in 🧵
Many asked, "How do I verify Gemini SFT works?" and "How do I prepare DPO data?"
With Vertex AI ENG team we published tutorials on preparing data and using custom metrics for evaluating tuning with Gemini 2.5 models.
Code in 🧵
Vertex AI has room from improvement with DevX. But when it comes to the heavy lifting like serving LLMs at scale to other Google AI platforms, we've got that nailed.
December 4, 2025 at 1:18 PM
Vertex AI has room from improvement with DevX. But when it comes to the heavy lifting like serving LLMs at scale to other Google AI platforms, we've got that nailed.
NEW VERTEX AI ENGINEERING BLOG: Scaling EAGLE-3 on Vertex AI with SGLang
We've launched an engineering blog to document our Vertex AI research. Our first post, in collaboration withSGLang, shares insights on implementing EAGLE-3 at scale.
Blog and code in the 🧵
We've launched an engineering blog to document our Vertex AI research. Our first post, in collaboration withSGLang, shares insights on implementing EAGLE-3 at scale.
Blog and code in the 🧵
December 2, 2025 at 2:30 PM
NEW VERTEX AI ENGINEERING BLOG: Scaling EAGLE-3 on Vertex AI with SGLang
We've launched an engineering blog to document our Vertex AI research. Our first post, in collaboration withSGLang, shares insights on implementing EAGLE-3 at scale.
Blog and code in the 🧵
We've launched an engineering blog to document our Vertex AI research. Our first post, in collaboration withSGLang, shares insights on implementing EAGLE-3 at scale.
Blog and code in the 🧵
🚀 NEW official docs for integrating ADK and A2A agents into Gemini Enterprise!
Docs in the 🧵
Docs in the 🧵
November 20, 2025 at 12:30 PM
🚀 NEW official docs for integrating ADK and A2A agents into Gemini Enterprise!
Docs in the 🧵
Docs in the 🧵
Vertex AI now supports Inline Source Deployment.
Deploy source code directly via API for:
> Git-native CI/CD
> No GCS buckets
> Auditable builds
Code and blog in the 🧵
Deploy source code directly via API for:
> Git-native CI/CD
> No GCS buckets
> Auditable builds
Code and blog in the 🧵
November 19, 2025 at 2:30 PM
Vertex AI now supports Inline Source Deployment.
Deploy source code directly via API for:
> Git-native CI/CD
> No GCS buckets
> Auditable builds
Code and blog in the 🧵
Deploy source code directly via API for:
> Git-native CI/CD
> No GCS buckets
> Auditable builds
Code and blog in the 🧵
If I were starting a new AI project today, I wouldn't default to standard frameworks.
We often stick to familiar tools for comfort, rather than exploring alternatives. The JAX AI stack and its components look exciting to me.
There's a lot to learn here: docs.jaxstack.ai/en/latest/
We often stick to familiar tools for comfort, rather than exploring alternatives. The JAX AI stack and its components look exciting to me.
There's a lot to learn here: docs.jaxstack.ai/en/latest/
November 15, 2025 at 7:08 AM
If I were starting a new AI project today, I wouldn't default to standard frameworks.
We often stick to familiar tools for comfort, rather than exploring alternatives. The JAX AI stack and its components look exciting to me.
There's a lot to learn here: docs.jaxstack.ai/en/latest/
We often stick to familiar tools for comfort, rather than exploring alternatives. The JAX AI stack and its components look exciting to me.
There's a lot to learn here: docs.jaxstack.ai/en/latest/
Supervised Fine-Tuning (SFT) adapts pre-trained models with labeled data but often misses aligning with user preferences.
Vertex AI now supports preference tuning (DPO) for Gemini 2.5 Flash and Flash-Lite, allowing you to use response pairs to adjust user preferences.
Code and docs in the 🧵
Vertex AI now supports preference tuning (DPO) for Gemini 2.5 Flash and Flash-Lite, allowing you to use response pairs to adjust user preferences.
Code and docs in the 🧵
November 14, 2025 at 6:00 AM
Supervised Fine-Tuning (SFT) adapts pre-trained models with labeled data but often misses aligning with user preferences.
Vertex AI now supports preference tuning (DPO) for Gemini 2.5 Flash and Flash-Lite, allowing you to use response pairs to adjust user preferences.
Code and docs in the 🧵
Vertex AI now supports preference tuning (DPO) for Gemini 2.5 Flash and Flash-Lite, allowing you to use response pairs to adjust user preferences.
Code and docs in the 🧵
Many have asked about agent memory and its differences from RAG.
We just released a whitepaper, "Context Engineering: Sessions and Memory," detailing how memory evolves from raw conversations to curated agent knowledge.
You can find the paper here: lnkd.in/euud4BUB
We just released a whitepaper, "Context Engineering: Sessions and Memory," detailing how memory evolves from raw conversations to curated agent knowledge.
You can find the paper here: lnkd.in/euud4BUB
November 12, 2025 at 10:11 AM
Many have asked about agent memory and its differences from RAG.
We just released a whitepaper, "Context Engineering: Sessions and Memory," detailing how memory evolves from raw conversations to curated agent knowledge.
You can find the paper here: lnkd.in/euud4BUB
We just released a whitepaper, "Context Engineering: Sessions and Memory," detailing how memory evolves from raw conversations to curated agent knowledge.
You can find the paper here: lnkd.in/euud4BUB
TIME to upgrade: ADK introduces Visual Agent Builder 🚀
Last week, in the latest ADK release (v1.18.0), the team introduced a low-code Visual Agent Builder, along with new observability and testing features.
Release notes and a blog from Thomas Chong about the visual builder in the 🧵
Last week, in the latest ADK release (v1.18.0), the team introduced a low-code Visual Agent Builder, along with new observability and testing features.
Release notes and a blog from Thomas Chong about the visual builder in the 🧵
November 10, 2025 at 7:00 PM
TIME to upgrade: ADK introduces Visual Agent Builder 🚀
Last week, in the latest ADK release (v1.18.0), the team introduced a low-code Visual Agent Builder, along with new observability and testing features.
Release notes and a blog from Thomas Chong about the visual builder in the 🧵
Last week, in the latest ADK release (v1.18.0), the team introduced a low-code Visual Agent Builder, along with new observability and testing features.
Release notes and a blog from Thomas Chong about the visual builder in the 🧵
Vertex AI Agent Engine adds Memory Revisions!
Trusting an agent's memory is tricky. Is chat info verified? Without history, it's guess. Memory Revisions (preview) helps with version control through snapshots for each change.
Code & doc in 🧵
Trusting an agent's memory is tricky. Is chat info verified? Without history, it's guess. Memory Revisions (preview) helps with version control through snapshots for each change.
Code & doc in 🧵
November 7, 2025 at 4:00 PM
Vertex AI Agent Engine adds Memory Revisions!
Trusting an agent's memory is tricky. Is chat info verified? Without history, it's guess. Memory Revisions (preview) helps with version control through snapshots for each change.
Code & doc in 🧵
Trusting an agent's memory is tricky. Is chat info verified? Without history, it's guess. Memory Revisions (preview) helps with version control through snapshots for each change.
Code & doc in 🧵
🚢 The New Vertex AI Agent Builder is OUT!
Vertex AI launched major updates to Vertex AI Agent Builder for easier deployment and scaling agents in production.
I'm working with the Agent Engine team to release content on our new features, starting today with Memory Bank. Stay tuned!
Vertex AI launched major updates to Vertex AI Agent Builder for easier deployment and scaling agents in production.
I'm working with the Agent Engine team to release content on our new features, starting today with Memory Bank. Stay tuned!
November 6, 2025 at 5:51 PM
🚢 The New Vertex AI Agent Builder is OUT!
Vertex AI launched major updates to Vertex AI Agent Builder for easier deployment and scaling agents in production.
I'm working with the Agent Engine team to release content on our new features, starting today with Memory Bank. Stay tuned!
Vertex AI launched major updates to Vertex AI Agent Builder for easier deployment and scaling agents in production.
I'm working with the Agent Engine team to release content on our new features, starting today with Memory Bank. Stay tuned!
Spent some time last week benchmarking LLMs on Vertex AI. I couldn't find a tutorial on using the @vllm_project bench library with Vertex, so I made one.
It's a walkthrough that compares Llama 4 (baseline vs. EAGLE) on 8x H100s and includes the code patch needed to make vLLM work.
Hope it helps! 👇
It's a walkthrough that compares Llama 4 (baseline vs. EAGLE) on 8x H100s and includes the code patch needed to make vLLM work.
Hope it helps! 👇
November 3, 2025 at 4:30 PM
Spent some time last week benchmarking LLMs on Vertex AI. I couldn't find a tutorial on using the @vllm_project bench library with Vertex, so I made one.
It's a walkthrough that compares Llama 4 (baseline vs. EAGLE) on 8x H100s and includes the code patch needed to make vLLM work.
Hope it helps! 👇
It's a walkthrough that compares Llama 4 (baseline vs. EAGLE) on 8x H100s and includes the code patch needed to make vLLM work.
Hope it helps! 👇
🔥 Benchmarking a new optimization integrated by the Model Garden team for serving LLMs on Vertex AI.
Can't wait to share!
#VertexAI #LLMs #Benchmarking #Optimization #ModelGarden #LLMServing
Can't wait to share!
#VertexAI #LLMs #Benchmarking #Optimization #ModelGarden #LLMServing
October 30, 2025 at 5:41 AM
🔥 Benchmarking a new optimization integrated by the Model Garden team for serving LLMs on Vertex AI.
Can't wait to share!
#VertexAI #LLMs #Benchmarking #Optimization #ModelGarden #LLMServing
Can't wait to share!
#VertexAI #LLMs #Benchmarking #Optimization #ModelGarden #LLMServing
Another week, another episode of the Agent Factory podcast!
This time, Amit and I welcomed a special guest: Ravin Kumar from Google DeepMind. He shared insights on building open models with agentic capabilities.
Stay tuned! The episode will soon be on the Google Cloud Tech YouTube channel👇
This time, Amit and I welcomed a special guest: Ravin Kumar from Google DeepMind. He shared insights on building open models with agentic capabilities.
Stay tuned! The episode will soon be on the Google Cloud Tech YouTube channel👇
October 28, 2025 at 9:00 PM
Another week, another episode of the Agent Factory podcast!
This time, Amit and I welcomed a special guest: Ravin Kumar from Google DeepMind. He shared insights on building open models with agentic capabilities.
Stay tuned! The episode will soon be on the Google Cloud Tech YouTube channel👇
This time, Amit and I welcomed a special guest: Ravin Kumar from Google DeepMind. He shared insights on building open models with agentic capabilities.
Stay tuned! The episode will soon be on the Google Cloud Tech YouTube channel👇
Since the beginning of the year, I've wanted to dedicate time on some content about LLM inference in a way that was accessible to everyone. Today I'm excited to launch the 1st learning path on LLM inference, built with NVIDIA!
Check out the course and blog in the 🧵 And stay tuned...more is coming!
Check out the course and blog in the 🧵 And stay tuned...more is coming!
October 27, 2025 at 3:33 PM
Since the beginning of the year, I've wanted to dedicate time on some content about LLM inference in a way that was accessible to everyone. Today I'm excited to launch the 1st learning path on LLM inference, built with NVIDIA!
Check out the course and blog in the 🧵 And stay tuned...more is coming!
Check out the course and blog in the 🧵 And stay tuned...more is coming!
ADK just released v1.17.0!
The team rolled out features focused on secure code execution, stateful debugging, and better DevX.
Check out the full release notes in 🧵. And keep an eye on the repo for the next ADK community call!
The team rolled out features focused on secure code execution, stateful debugging, and better DevX.
Check out the full release notes in 🧵. And keep an eye on the repo for the next ADK community call!
October 22, 2025 at 9:35 PM
ADK just released v1.17.0!
The team rolled out features focused on secure code execution, stateful debugging, and better DevX.
Check out the full release notes in 🧵. And keep an eye on the repo for the next ADK community call!
The team rolled out features focused on secure code execution, stateful debugging, and better DevX.
Check out the full release notes in 🧵. And keep an eye on the repo for the next ADK community call!
In the upcoming webinar, together with Alex Notov from @AnthropicAI we're building a complete multi-agent system, exploring the key protocols (MCP & A2A) and how to scale agents using Claude on Vertex AI Agent Engine.
RVSP and tutorial in the 🧵
RVSP and tutorial in the 🧵
October 21, 2025 at 4:30 PM
In the upcoming webinar, together with Alex Notov from @AnthropicAI we're building a complete multi-agent system, exploring the key protocols (MCP & A2A) and how to scale agents using Claude on Vertex AI Agent Engine.
RVSP and tutorial in the 🧵
RVSP and tutorial in the 🧵
Vertex AI Agent Engine is in the Cloud Foundation Fabric!
We launched a Terraform module for agent deployment on Vertex AI, initially needing a local Python script and GCS uploads.
The new blueprint automates agent serialization and packaging during the Terraform apply cycle.
Link in the 🧵
We launched a Terraform module for agent deployment on Vertex AI, initially needing a local Python script and GCS uploads.
The new blueprint automates agent serialization and packaging during the Terraform apply cycle.
Link in the 🧵
October 20, 2025 at 4:19 PM
Vertex AI Agent Engine is in the Cloud Foundation Fabric!
We launched a Terraform module for agent deployment on Vertex AI, initially needing a local Python script and GCS uploads.
The new blueprint automates agent serialization and packaging during the Terraform apply cycle.
Link in the 🧵
We launched a Terraform module for agent deployment on Vertex AI, initially needing a local Python script and GCS uploads.
The new blueprint automates agent serialization and packaging during the Terraform apply cycle.
Link in the 🧵
This is why I'm excited about Google Cloud's agent builder stack!
Building multi-agent systems using various models and frameworks can be challenging. So today I spent some time on what you can build and deploy with ADK, MCP, A2A, Agent Engine, and Vertex AI.
Full tutorial dropping soon!
Building multi-agent systems using various models and frameworks can be challenging. So today I spent some time on what you can build and deploy with ADK, MCP, A2A, Agent Engine, and Vertex AI.
Full tutorial dropping soon!
October 19, 2025 at 6:57 PM
This is why I'm excited about Google Cloud's agent builder stack!
Building multi-agent systems using various models and frameworks can be challenging. So today I spent some time on what you can build and deploy with ADK, MCP, A2A, Agent Engine, and Vertex AI.
Full tutorial dropping soon!
Building multi-agent systems using various models and frameworks can be challenging. So today I spent some time on what you can build and deploy with ADK, MCP, A2A, Agent Engine, and Vertex AI.
Full tutorial dropping soon!
🚀 vLLM on TPU just got a massive upgrade!
Google and vLLM announced a new backend uses tpu-inference for efficient PyTorch and JAX models on TPUs.
Check out the full blog and try it on Vertex AI with the new vLLM TPU container!
Google and vLLM announced a new backend uses tpu-inference for efficient PyTorch and JAX models on TPUs.
Check out the full blog and try it on Vertex AI with the new vLLM TPU container!
October 16, 2025 at 9:00 PM
🚀 vLLM on TPU just got a massive upgrade!
Google and vLLM announced a new backend uses tpu-inference for efficient PyTorch and JAX models on TPUs.
Check out the full blog and try it on Vertex AI with the new vLLM TPU container!
Google and vLLM announced a new backend uses tpu-inference for efficient PyTorch and JAX models on TPUs.
Check out the full blog and try it on Vertex AI with the new vLLM TPU container!
This morning, I checked out Vertex AI docs and was impressed by the open-source models as APIs.
Model as a Service (MaaS) gives access to large open models via a managed, serverless API, removing the need for your infrastructure.
Check the new documentation in 🧵 to learn more
Model as a Service (MaaS) gives access to large open models via a managed, serverless API, removing the need for your infrastructure.
Check the new documentation in 🧵 to learn more
October 16, 2025 at 5:30 PM
This morning, I checked out Vertex AI docs and was impressed by the open-source models as APIs.
Model as a Service (MaaS) gives access to large open models via a managed, serverless API, removing the need for your infrastructure.
Check the new documentation in 🧵 to learn more
Model as a Service (MaaS) gives access to large open models via a managed, serverless API, removing the need for your infrastructure.
Check the new documentation in 🧵 to learn more
Can you use Agent Engine services on GKE or Cloud Run? Yes! You can combine managed services like memory bank with your preferred runtime.
Check out the two new tutorials in 🧵 on building AI agents using the Agent Development Kit (ADK) + Vertex AI Agent Engine for Sessions & Memory.
Check out the two new tutorials in 🧵 on building AI agents using the Agent Development Kit (ADK) + Vertex AI Agent Engine for Sessions & Memory.
October 16, 2025 at 3:00 PM
Can you use Agent Engine services on GKE or Cloud Run? Yes! You can combine managed services like memory bank with your preferred runtime.
Check out the two new tutorials in 🧵 on building AI agents using the Agent Development Kit (ADK) + Vertex AI Agent Engine for Sessions & Memory.
Check out the two new tutorials in 🧵 on building AI agents using the Agent Development Kit (ADK) + Vertex AI Agent Engine for Sessions & Memory.
🚀 Deploying agents on Vertex AI Agent Engine with Terraform!
Vertex AI launched a Terraform resource to deploy agents using custom classes or agentic frameworks like ADK.
Check out the notebook and blog post for the full code and a step-by-step guide in the 🧵
Vertex AI launched a Terraform resource to deploy agents using custom classes or agentic frameworks like ADK.
Check out the notebook and blog post for the full code and a step-by-step guide in the 🧵
October 15, 2025 at 3:00 PM
🚀 Deploying agents on Vertex AI Agent Engine with Terraform!
Vertex AI launched a Terraform resource to deploy agents using custom classes or agentic frameworks like ADK.
Check out the notebook and blog post for the full code and a step-by-step guide in the 🧵
Vertex AI launched a Terraform resource to deploy agents using custom classes or agentic frameworks like ADK.
Check out the notebook and blog post for the full code and a step-by-step guide in the 🧵
🗓️ SAVE THE DATE: ADK Community Call #1 is coming!
Join us next week for the 1st ADK community call! In this 1-hour session, we'll share the technical roadmap, address technical questions and discuss contributions.
🗓️ Date: October 15, 2025
⏰ Time: 9:30 AM - 10:30 AM PST
🔗 Virtual (links in 🧵)
Join us next week for the 1st ADK community call! In this 1-hour session, we'll share the technical roadmap, address technical questions and discuss contributions.
🗓️ Date: October 15, 2025
⏰ Time: 9:30 AM - 10:30 AM PST
🔗 Virtual (links in 🧵)
October 10, 2025 at 5:21 PM
🗓️ SAVE THE DATE: ADK Community Call #1 is coming!
Join us next week for the 1st ADK community call! In this 1-hour session, we'll share the technical roadmap, address technical questions and discuss contributions.
🗓️ Date: October 15, 2025
⏰ Time: 9:30 AM - 10:30 AM PST
🔗 Virtual (links in 🧵)
Join us next week for the 1st ADK community call! In this 1-hour session, we'll share the technical roadmap, address technical questions and discuss contributions.
🗓️ Date: October 15, 2025
⏰ Time: 9:30 AM - 10:30 AM PST
🔗 Virtual (links in 🧵)
🚀 Deploying open models with Terraform on Vertex AI!
Vertex AI Model Garden just launched the google_vertex_ai_endpoint_with_model_garden_deployment Terraform resource to manage your open model deployment (Hugging Face or Model Garden) with one unique main.tf.
Docs and code in the 🧵
Vertex AI Model Garden just launched the google_vertex_ai_endpoint_with_model_garden_deployment Terraform resource to manage your open model deployment (Hugging Face or Model Garden) with one unique main.tf.
Docs and code in the 🧵
October 8, 2025 at 3:30 PM
🚀 Deploying open models with Terraform on Vertex AI!
Vertex AI Model Garden just launched the google_vertex_ai_endpoint_with_model_garden_deployment Terraform resource to manage your open model deployment (Hugging Face or Model Garden) with one unique main.tf.
Docs and code in the 🧵
Vertex AI Model Garden just launched the google_vertex_ai_endpoint_with_model_garden_deployment Terraform resource to manage your open model deployment (Hugging Face or Model Garden) with one unique main.tf.
Docs and code in the 🧵