An Open Letter from

There are too many jargons and buzzwords in the DApp space keeping the community from fully understanding what’s really going on.

I am writing this open letter, to answer the most common queries we’ve received and explain our platform in the simplest way possible. Let’s begin -

:round_pushpin:What is ELNA ?

It is a decentralized AI agent-creation platform.

Let me make that simple -

It’s a platform that can help you create applications like:

  • Personal finance manager
  • Social media manager
  • Personal tutor
  • Telemedicine screening agents for doctors
  • Legal assistants for lawyers
  • Data analyst with real-time intel

The possibilities are truly endless.

:round_pushpin:Is ELNA a chatbot ?

Not really. Sure, you can make a chatbot using ELNA, but it’s only the tip of the iceberg of what ELNA is capable of.

ELNA provides you with the framework, tools and integration to create autonomous agents using any LLM - these agents can be designed to execute tasks and operations like an average human.

:round_pushpin:Why ELNA?

Let me explain this with an example -

Think of ELNA as a workshop - we have all the lathes, tools and machineries, along with a set of highly skilled workforce. Let’s say you’re trying to build a mini airplane on your own. It’s gonna be a tonne of effort - you’re going to have to source the right propeller, fuselage, wings, engine and more, eventually integrating the parts together to create a fully functional machine and run it, all by yourself.

Instead, you come to us and tell us your requirement, choose the features you need. And sit tight. We create the machine for you, solve your problem, and support in running the same moving forward.

In this workshop analogy, the primary tool is the LLM. Like a versatile engine that can power a tractor, scooter, or airplane - the same way an LLM can power a variety of applications, from support agents, to automated trading and even marketing solutions. ELNA thus serves as the facilitator, offering the tools to craft AI agent applications of your choice, with the least amount of effort from your end.


:white_small_square:We create:

Custom Autonomous AI Agents to solve your specific problem statements.


On- and off-chain LLMs

:white_small_square:By providing our:

(1) Vector database, (2) prompt engine, (3) integration tools to connect to platforms like X, Instagram, Wikipedia et al, (4) Tools to bring the LLMs on-chain, and fine-tune them.

What I want to make clear is: ELNA creates and customises an agent to your specific needs, provide prompts, add knowledge base to raw LLMs and execute integration capabilities like connecting to twitter, telegram or perform mathematical calculations, analytics with real-time data etc., helping you bring life to your dream agents in supporting your customers/self.

:round_pushpin:What ELNA does not do

We do not create LLMs, because LLMs are not agents. They’re only one of the building blocks in the toolkit required to create autonomous AI agents. We are capable of bringing all LLMs on-chain and fine-tuning them. A whole bunch of other resources, as stated above, are necessary to give an LLM a full-fledged use case, which ELNA provides.


I am not a developer or AI expert, but I see great potential in “AI on IC” and am looking into DeAI’s working group and kinic’s team. I believe ELNA could be the future for trustless agents which execute on chain.

To do it all on-chain:

  • Upload and host pre-trained LLMs
  • Inference
  • Fine-tuning of the pre-trained LLMs (as the case may be)

The above is my understanding of what is needed.
To what extent is the current state of ELNA’s AI agents executable on-chain? Does you need GPU subnets for AI agent?

According to the ELNA diagram, it looks like “Training Canisters” is used to instruct fine-tuning in PEFT & QLoRA based on “User Data” in Hugging Face + SageMaker. Also, is the “LLM Base Model” stored in the “wasm-native stable memory” the entire pre-trained LLM or a part of the fine-tuned LLM?
Please correct me if there are any mistakes.

1 Like

We would appreciate it if you could answer us if possible to promote understanding.


Hi @hokosugi ,
Thank you for expressing your interest and sharing your insights on ELNA’s . I appreciate your attention to the details of our project.

I will be writing a detail post explaining how its done and driving more into technical aspects.

Now in short .
You need to have the infra that include the inference engines, framework etc to be built in web-assembly that need to be compatible with IC canisters.
Right now due to instruction/ memory constrains and more importantly GPU we have limitations on running larger infra/models. At ELNA we are taking a hybrid system

We have high hopes that Dfinity with take the right steps.

We are building an infrastructure to build LLM agents (that includes prompt engine, vector db and other tool integrations) where you can connect your LLM (On-chain & Off-chain ) to built your custom autonomous agents.

and in future we do have plans for fine-tuning llms on chain and other cool stuffs ones the GPU canisters are live

1 Like

Thank you for your response.
It is true that there is a limit to the performance of the current IC, and I think it is the right approach to present a hybrid product first, even if it is criticized.
I think autonomous AI agents could be a killer app considering the privacy potential of ICs!
I’m waiting for more details to be uploaded.