GSoC 2026: Project 13 - No-Code AI Workflow Automation with n8n and OpenVINO Model Server for GPU/NPU

[Nandkishore-04]

Hi @praveenkk123, @mjdomeik

I'm Nand Kishore, a final year student pursuing my BTech CSE in (AI and Data Science), and project idea 13 really caught my attention. I’ve been exploring the OpenVINO GenAI codebase and currently have two active PRs under review. I have recently been working on building n8n workflows for Businesses and reducing the manual work.

My understanding of the project: Build a n8n community nodes that let non-technical users run AI inference on Intel hardware (CPU/GPU/NPU) through OVMS — selecting models, choosing devices, and running inference all from the n8n UI.

I built a working prototype to explore the integration approach,
ARCHITECTURE:
n8n (custom node) → Gateway (tokenizes text) → OVMS (runs inference) → Results back to n8n

IMPLEMENTATION DETAILS:

1. Custom n8n community node (TypeScript) with 3 operations: list models, check status, run inference
2. Gateway service (Python) handles tokenization so n8n users send plain text instead of raw tensors
3. DistilBERT sentiment classifier converted to OpenVINO IR format
4. Full Docker Compose setup — one command to run everything

Repo: https://github.com/Nandkishore-04/n8n-openvino

OpenVINO GenAI Contributions
#3337 — Node.js N-API bindings for Text2VideoPipeline
#3155 — Introduced C API for Text-to-Speech pipeline and SpeechT5 (Fixes #2302)

Looking for Feedback On:

1. Is the Gateway approach (tokenization layer between n8n and OVMS) the right architecture?
2. Should the n8n node support gRPC alongside REST for better performance?

I’d appreciate feedback on whether this architecture aligns with the intended scope and where you see the highest priority.

regards,
Nand Kishore

[praveenkk123]

Hi Nand Kishore,
Thank you for sharing this—great progress so far, and the prototype looks like a promising foundation.

Can you please prioritize a full agentic workflow using n8n + OVMS locally, with a model that supports agent-style interactions and function/tool calling? I think that would help align strongly with the intended direction of the project.

On your architecture questions:

The Gateway approach seems reasonable for now, as long as it continues to work well for agentic/tool-calling workflows.
For the moment, REST is sufficient; no need to add gRPC yet unless you later see a clear performance need.

As part of the next steps, it would be great if you could implement:

1. A working local agentic flow,
2. Any node-level updates needed for tool/function-calling behavior, and
3. Brief notes on reliability and how the agent works

Please let me know if you got any further questions

Thanks
Praveen

[Nandkishore-04]

Thank you for your feedback @praveenkk123 , this is really helpful. I will implement the agentic workflow with function/tool calling, and update the repo, share the progress once it's done. Thank you! 😄

[Nandkishore-04]

Hi @praveenkk123,
I've implemented what you suggested and wanted to share what I've built:

What's working:

Custom n8n node that connects to OpenVINO Model Server

1. Classic inference: DistilBERT sentiment analysis via OVMS + Gateway
2. Agentic workflow: Qwen2.5-1.5B on OVMS acts as a customer support triage agent — it calls tools like sentiment analysis, Knowledge Base lookup, and ticket creation based on what it decides at runtime
3. The agent takes different paths depending on sentiment (2-4 tool calls)
4. Full Docker Compose setup (OVMS, OVMS-LLM, Gateway, n8n, PostgreSQL)
5. Screenshots and sample outputs are in the README file.

Repo: https://github.com/Nandkishore-04/n8n-openvino

I'd appreciate any feedback on this, and would love your guidance on shaping a strong GSoC proposal for this project.
Thanks,
Nand Kishore

[praveenkk123]

Hi Nand Kishore,
This looks great and thanks for the quick turnaround on this. I have a question — I see in your screenshots you implemented the Agents, which is great, but I don't see you using any specific n8n built-in AI Agentic nodes. Can you also see if you can implement the "AI Agent" n8n node where you can add memory, call custom tools or pre-built tools, and also add a complete workflow as a tool to your agent? Can you explore that path, please?

Thanks
Praveen

[Nandkishore-04]

sure, just to confirm-you want me to use n8n's built-in AI Agent node with OVMS as the LLM backend, and show it working with memory, custom tools (like my OVMS sentiment node), pre-built tools, and a workflow as a tool. Should I keep the existing custom agent demo alongside this, or replace it?

[praveenkk123]

Lets keep your existing custom agent demo as its interesting. Please see if you can create an use case that combines both of these together in a single flow. If it is not feasible we can create a new workflow with AI agent

Thanks
Praveen

[Nandkishore-04]

Thanks, I'll make a workflow where I will combine both of these(custom node and pre-built AI agent node) for an actual use case. I'll explore this and update you once it's done!

[Nandkishore-04]

Hi @praveenkk123 ,

I kept the existing custom agent demo as is and created a combined workflow that uses both together in a single flow:

Combined Workflow (/support-v2):

Custom OVMS Node runs DistilBERT sentiment analysis (fast, deterministic)
n8n Built-in AI Agent receives the sentiment result and handles reasoning with:

1. OpenAI Chat Model → Qwen2.5-1.5B on OVMS (via gateway /v1 proxy)
2. Window Buffer Memory
3. Knowledge Base Lookup (Code Tool)
4. Create Ticket (Code Tool)
5. Calculator (pre-built tool)
6. Sentiment Analysis (Workflow-as-a-Tool → sub-workflow using the custom OVMS node)

So both the custom node and n8n's built-in AI Agent work together — the custom node handles classification, the AI Agent handles reasoning and tool orchestration.

One limitation: Qwen2.5-1.5B is too small to reliably call tools through the built-in AI Agent. It tends to generate text describing the tool calls instead of executing them. The custom agent with the nudge system handles this correctly. I've documented this with a comparison table in the repo. A larger model (7B+) would resolve this.

Everything is pushed with screenshots and docs: https://github.com/Nandkishore-04/n8n-openvino (please refer the combined workflow demo screenshots )

Let me know if you'd like me to take a different approach .

[praveenkk123]

Hi Nand Kishore, This looks great. Let me get back to you on this on the next steps.
Thanks
Praveen

[Nandkishore-04]

Thanks, Praveen! Looking forward to it.

[Nandkishore-04]

Hi @praveenkk123,
In the meantime, I’ll begin drafting the formal proposal based on our current discussion so I can share a version with you for feedback soon.

Thanks
Nand kishore

[praveenkk123]

Hi Nand Kishore,
Thanks. When you are going to submit the proposal please make sure you include advanced and additional features to this workflow. Do not implement any of the below new features now until your project gets selected and the official coding starts. Now that you did some basic AI Agentic workflows you can implement advanced features like HTTP requests, web hooks, adding a complete workflow as a tool to AI Agent, write custom code node as a tool and error triggers etc. You can investigate to add more features if you feel like they fit in your workflow
Thanks
Praveen

[naraen-ram]

Hi @praveenkk123,

When you have time please go through my open discussion regarding project 13. I would really appreciate your input before I start drafting my formal proposal.

Discussion link: #34710

Thanks,
Naraen Rammoorthi

[Nandkishore-04]

Hi Praveen,

Thank you for the feedback. I'll include the advanced features in my proposal — HTTP requests, workflow-as-a-tool, custom code tools, error triggers, and any additional features that fit the workflow.

I'll draft the proposal and share it with you for review before submitting.

Thanks,
Nand Kishore

[Lokpradeep-creater]

Hi mentors, I'm Lok Pradeep, a CSE student at NIAT chennai. I have hands-on experience building AI workflows with n8n (recently built an AI News Summarizer) and working with voice AI/LLMs. I am highly interested in contributing to Project #13 (n8n Workflow Automation) and would love to request guidance on a prerequisite task to prove my skills. Looking forward to contributing!

[Nandkishore-04]

Hey @praveenkk123,
I've drafted my GSoC proposal and shared it with you via praveen.k.kundurthy@intel.com. I would really appreciate your feedback before I submit it.

Thanks,
Nand Kishore

[mjdomeik]

Nand, Can you share with me as well? ***@***.******@***.***> Praveen is out this week. Thanks! Max From: Nand kishore ***@***.***> Sent: Wednesday, March 25, 2026 6:17 AM To: openvinotoolkit/openvino ***@***.***> Cc: Domeika, Max J ***@***.***>; Mention ***@***.***> Subject: Re: [openvinotoolkit/openvino] GSoC 2026: Project 13 - No-Code AI Workflow Automation with n8n and OpenVINO Model Server for GPU/NPU (Discussion #34422) Hey @praveenkk123<https://github.com/praveenkk123>, I've drafted my GSoC proposal and shared it with you via ***@***.******@***.***>. I would really appreciate your feedback before I submit it. Thanks, Nand Kishore — Reply to this email directly, view it on GitHub<#34422?email_source=notifications&email_token=AAKC5ZKOU6DHJ4VKQUIPIS34SPL4TA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNRTGA4TONRTUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-16309763>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AAKC5ZKIHRJ7OTP7XR7VZ7T4SPL4TAVCNFSM6AAAAACWDOUK5SVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMMZQHE3TMMY>. You are receiving this because you were mentioned.Message ID: ***@***.******@***.***>>

[Nandkishore-04]

Hi @mjdomeik ,

Thanks for reaching out! Could you share your email so I can give you access to my proposal?
In the meantime, here's the GitHub repo with the pre-proposal demo: https://github.com/Nandkishore-04/n8n-openvino

Thanks,
Nand Kishore

[mjdomeik]

***@***.******@***.***> Thanks! Max From: Nand kishore ***@***.***> Sent: Wednesday, March 25, 2026 10:49 AM To: openvinotoolkit/openvino ***@***.***> Cc: Domeika, Max J ***@***.***>; Mention ***@***.***> Subject: Re: [openvinotoolkit/openvino] GSoC 2026: Project 13 - No-Code AI Workflow Automation with n8n and OpenVINO Model Server for GPU/NPU (Discussion #34422) Hi @mjdomeik<https://github.com/mjdomeik> , Thanks for reaching out! Could you share your email so I can give you access to my proposal? In the meantime, here's the GitHub repo with the pre-proposal demo: https://github.com/Nandkishore-04/n8n-openvino Thanks, Nand Kishore — Reply to this email directly, view it on GitHub<#34422?email_source=notifications&email_token=AAKC5ZLUBWKPHVGB45RVURD4SQLYDA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNRTGEZTAOJTUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-16313093>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AAKC5ZOHVAG46Q6BG77YKSL4SQLYDAVCNFSM6AAAAACWDOUK5SVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMMZRGMYDSMY>. You are receiving this because you were mentioned.Message ID: ***@***.******@***.***>>

[Nandkishore-04]

Hi @mjdomeik,

I've sent the proposal draft to max.j.domeika@intel.com — please let me know if that's the correct email, or if I should resend it to a different one.

Also, just a heads up - your email appeared masked in the GitHub notification I received (showing as @.***), which is why I couldn't see it directly. GitHub automatically hides email addresses in notifications as a privacy measure.

You can also reach me directly at nkk01414@gmail.com.

Thanks,
Nand Kishore